Ed Zitron's Where's Your Ed At Yesterday

The OpenAI Bubble

Thanks for reading this week’s free Where’s Your Ed At newsletter. As I said last week, I’m taking the rest of this week off, so there won’t be a premium on Friday. That said, if you aren’t already a member, now’s a great time to subscribe.  To celebrate the one year anniversary of the premium newsletter, I’m offering a sale on one-year subscriptions. Between now and midnight July 22, you can get a permanent annual rate of just $60— a $10 discount on the usual price of $70 - for life. Click here for the offer . In addition to getting access to the entire back catalog of premium posts, you’ll also receive one additional post each week — usually anywhere between 10,000 and 20,000 words — covering the most pressing topics in the AI bubble - the best value in tech analysis. Highlights include last week's Hater's Guide To The Memory Crisis - a guide to how AI made everything more expensive - How OpenAI Kills Oracle (which pairs nicely with the Hater's Guide To Oracle ), The Hater's Guide To NVIDIA , The Hater's Guides To Private Credit and Private Equity , and how the entire AI Compute Demand Story Is A Lie . Today’s piece is one of the largest free newsletters I’ve ever written, and pulls together the last six months of my work. And it all starts with a question: how much do you trust Sam Altman? The stock market and (to some extent) the global economy rests on your answer. You see, OpenAI has become one of the largest liabilities in recent economic history. You can argue that OpenAI’s no longer the focal point of the AI bubble — you can talk all you want about open source models or Anthropic or any number of other elements — but without OpenAI, the AI industry doesn’t exist, and the justification for trillions of dollars of capex evaporates.  The AI bubble isn’t a result of any actual return on investment — whether that be in purely monetary terms, like revenue or profitability , productivity gains, or anything tangible or measurable. Rather, it’s an episode of cult-like psychosis that infected the brains of some of the most powerful and wealthy individuals and institutions, where the powerful mythology of a company inspired — and been used to inspire — the greatest capital misallocation in history.  As much as this’ll piss some people off, I fully believe that the only reason this has kept going so long is that OpenAI has yet to collapse. Its failure would be a watershed moment — the Lehman Brothers of the AI bubble, and an event that would define the end of one epoch, the start of another, and that would shake the afflicted out of that psychosis. Absent this wake-up call, NVIDIA has continued to sell GPUs, the coffers of the semiconductor industry have continued to swell, and more and more spending commitments have been made.  Look. OpenAI intends to burn over $852 billion by the end of 2030 . It accounts for $748 billion of the remaining performance obligations of Microsoft, Amazon, and Oracle, on top of at least another $70 billion of RPOs across Cerebras , CoreWeave , Nebius, IREN, Lambda, and Nscale (per Kakashii), and plans to spend indeterminate billions’-worth of Broadcom “Jalapeno” chips . It intends to spend $50 billion or more on compute this year , which I estimate is more than 50% of all global AI compute spend (with OpenAI taking up 50%+ of all AI compute infrastructure ).  OpenAI can only afford to pay that as a result of its latest (assuming it fully closes) $122 billion funding round , of which it has received at least $50 billion, with $20 billion from SoftBank (of $30 billion, with the third tranche due October 1, 2026 ). NVIDIA mentioned in its latest quarterly earnings report that it “estimate[d] that one AI research and deployment company contributed to a meaningful amount of [its] revenue by purchasing cloud services from [its] customers in the first quarter of fiscal year 2027,” referring, of course, to OpenAI. OpenAI is the reason anyone cares about AI. In March 2019 ( per JustDario ), NVIDIA bought a company called Mellanox that made the high-speed networking tech necessary to create AI GPU clusters, and four months after that, Microsoft invested a billion dollars in OpenAI and started buying AI GPUs and building AI infrastructure for it. By March 2020, NVIDIA would ship its A100 GPU , and in May 2020 , Microsoft would announce it had built a supercomputer just for OpenAI with “more than 285,000 CPU cores [and] 10,000 GPUs.” The launch of ChatGPT in November 2022 came at the perfect time for a tech industry that had run out of ideas and was flirting with a prolonged depression. The IPO market had collapsed , interest hikes killed the Zero Interest Free era dead, pandemic era overhiring began to unwind with some of the worst layoffs in the history of the industry , global venture funding dwindled after historic overinvestment in 2021 , and tech stocks took a massive beating .  For the first time, the tech industry was forced to cut its cloth in accordance with its means — something which it has historically been loath to do. Big tech was unpopular, both with investors and the general public. The excesses of the past decade — combined with the growing frustration with, for lack of a better word, “tech exceptionalism,” where it believed that the rules which governed the rest of the world didn’t apply to Silicon Valley — had tested the patience of both regulators and lawmakers. And, in the absence of “one more thing” — a big, splashy, game-changing product category — it no longer had an excuse for its prodigal spending, or its regular breaking of the rules, both written and unwritten, that govern society. The existence of OpenAI justified an era of mania and opulence. Hyperscalers, bereft of new hypergrowth ideas , were able to point at the fact that ChatGPT had “the fastest growing userbase of all time” and the Microsoft “supercomputer” that built it and tell their investors that if they didn’t invest, they’d be left behind , with Amazon , Meta , and Google announcing their own nebulous “supercomputers” in 2023.  By the end of 2023, NVIDIA had sold 500,000 A100 GPUs , and the only reason it did so was because of ChatGPT’s rapid growth . Sam Altman’s brief ouster only sought to inflate the AI bubble by adding a layer of dull palace intrigue to a tech industry bereft of whimsy or character — and helped further entrench Microsoft’s role as the paternalistic benefactor of OpenAI, which made sure that Altman returned to the helm . To be clear, when I say “rapid growth,” I mean that OpenAI hit 100 million weekly active users by the end of 2023 and had about $108 million in monthly revenue . Microsoft would invest $10 billion more that year , with the majority of that funding coming in the form of credits to be used on Microsoft Azure . OpenAI is also the reason that Anthropic exists — not just because multiple founders came from the company, but because both Google and Amazon both agreed to give it a total of $6 billion in 2023 as a means of “competing” with Microsoft’s new obsession, which allowed both to justify spending further hundreds of billions of dollars “to make sure they didn’t miss out on AI.”  When you remove the term “AI” from the equation, this all seems a little ludicrous. $16 billion in equity investment on top of what was, by the end of 2023, over $150 billion in capital expenditures, all of which was pretty much justified by the fact that a single website had been very popular.  And the only reason either of these companies were able to grow was because of hyperscalers bankrolling their entire infrastructure.  In the fourth quarter of 2023 , global venture capital funding had dropped to its lowest levels since the third quarter of 2016, with American startups taking up $183.6 billion of the year’s investments. Venture capital alone couldn’t have — and wouldn’t have — actually backed OpenAI or Anthropic at the scale that was necessary to build their infrastructure, nor would there have been any of the hunger from hyperscalers or those providing debt for data centers without hyperscalers inflating both of these companies, almost entirely because of the success of OpenAI.   Remove OpenAI from the years 2020 through 2024 and the AI bubble wouldn’t have inflated at all. No other major AI companies showed any sign of life — not those peddled by hyperscalers, funded by venture capitalists, or those launched by other tech firms.  The only reason that any hyperscaler AI efforts have any revenue — and outside OpenAI and Anthropic it’s pretty meager! — is because they knew they could just sit there and keep saying “AI is the future” until their customers eventually gave in and tried it…largely because everybody was talking about ChatGPT .  Anthropic was considered an also-ran until early 2025, and only continued to get funded because people wanted to invest in the next OpenAI , and Anthropic’s initial funding rounds and infrastructure buildout were only justified in terms of competing with OpenAI.   Those $178.5 billion in US-based data center debt deals in 2025 ? Pretty much entirely justified by the growth of OpenAI and its rapacious hunger for compute, because outside of OpenAI (and eventually Anthropic), nobody else was using massive clusters of tens of thousands of GPUs, nor does a market for compute at that scale appeared to have popped up in the months and years since.  The largest consumers of compute remain Microsoft (for OpenAI), Google (for Anthropic), Amazon (for OpenAI and Anthropic), CoreWeave (for OpenAI and Anthropic), Meta (which is copying what the other hyperscalers are doing), and Oracle (for OpenAI). Otherwise, there’s very little evidence — and boy, have I looked — that there’s more than a few billion in demand for AI compute, and that’s being generous.  All of those investments — both in AI startups and data centers — existed to fund either the next OpenAI or become the next OpenAI’s landlord.  The assumption — because nobody ever thinks things through — was that because one OpenAI existed, many OpenAIs would bloom. That because one large customer of compute existed, the template had been built for future compute-intensive startups…and, again, because nobody ever thinks about anything, nobody ever stopped to realize that the reason there isn’t another OpenAI is because OpenAI and Anthropic are financial psy-ops by the largest software companies in the world.  The grim truth is that you can’t venture fund an AI lab. While OpenAI and Anthropic have raised nearly $300 billion in the last few years, their actual infrastructure costs — the GPUs and the data centers to power their services — were entirely funded by hyperscalers, likely costing another $250 billion in the process, given that Microsoft has said it spent $100 billion on its OpenAI relationship as of early 2026 .  Yet the real cost wasn’t just financial , but the experience and industrial know-how to actually execute on a massive infrastructure bailout. Other than Google, Microsoft, and Amazon, nobody else has the scale or experience to build the kind of AI clusters that OpenAI (and eventually Anthropic) needed.  We know that for a couple of reasons. First, because prior to 2023, there were few — if any — companies actually building AI computing clusters at the kind of scale demanded by OpenAI or Anthropic. The closest thing that one could point to were crypto-mining firms, and it’s telling that many of the neoclouds today (most famously Coreweave) started life running warehouses full of ASICs to mine Bitcoin and Ethereum. Second, because, based on conversations with people in the data center industry, the whole Overton window of what is considered to be a “big” facility has shifted. Previously, a 50MW data center would have been considered a significant (even noteworthy) development. These were the exception, and not the rule, with most data centers being vastly more modest affairs. The only companies which had any experience building at that scale were, for the most part, hyperscalers.   By treating OpenAI as a “venture backed startup,” hyperscalers created the illusion that this was the next type of big company that would in turn create the next great demand center in cloud computing , except the only reason that these companies existed was because of the hyperscalers themselves willing them into existence, funding them with incredible sums, and allowing them to burn as much money as they’d like.   This is why the idea that OpenAI will continue to grow infinitely is central to the mythology of the AI bubble. The existence of one OpenAI allows others to — no matter how illogical — imagine the existence of more OpenAIs, which in turn means that those OpenAIs will need just as much compute as OpenAI.   The dimwitted investor who believes this tripe can justify it through any number of different buy-side analysts or captured members of the media that talk about the “insatiable demand for compute,” pointing to capacity constraints ( caused by slow data center construction and — hah! — OpenAI and Anthropic taking up much of the world’s compute ) and increasing GPU prices as proof that actually, there’s tons of demand , all without ever really thinking too hard. The greatest trick that hyperscalers played was never backing down. By sinking more than a trillion dollars into AI capex without ever showing a single dollar of profit , they justified literally anyone investing in AI data centers under the logic that “the largest companies in the world couldn’t be wrong,” even if the reason they were doing so was to expand capacity for OpenAI and Anthropic, who the hyperscalers themselves incubated.   It is fundamentally illogical and insane for hyperscalers to have spent so much money on AI infrastructure, and the reason that few people will say so is because it was, until recently, considered radical to suggest that this was a waste of money, almost entirely because of the existence and continued growth of OpenAI. Whatever utility you may or may not get out of LLMs is irrelevant because it has not, for the most part, been what actually underpins data center investment. While accelerating gains in code generation (itself something that could have only happened without vast subsidies) might have helped grow Anthropic , the vast majority of data center capex has been built chasing the dragon of what AI could be rather than any connection to the revenues or economics of the companies at large — outside, of course, their compute spend.  This is the underlying greed that has driven this wasteful, reckless and destructive era — the belief that there will be another OpenAI and, as I’ve said, the chance to become the next OpenAI’s landlord. And because the media and analysts very rarely have original ideas, everybody justified (and justifies) the waste through the same tired mantras, saying it was “just like Uber ( nope !)” or “just like Amazon Web Services ( between 2003 and 2015, Amazon spent $29.7 billion on capex, normalized for inflation ).” And like any great investment bubble, the more money that piled in, the greater the fear of missing out, the more dollars that can be justified in turn, and the more-complex and deranged the mythology becomes, which is why you have noted venture capitalists claiming that AI labs have “90%+ inference margins,” a completely unproven statement that AI boosters cling to and repeat often enough that it’s taken as gospel, likely to avoid thinking about the fact that you can burn $14,000 in tokens on a $200-a-month ChatGPT subscription .  This kind of mythology only grows in an environment deliberately deprived of good information. The fact that we’re four years into this horrible bubble and still don’t have consistently-held consensus around the actual costs of large language models is a testament to an industry-wide effort to suppress them.  OpenAI, Anthropic, Microsoft, Google, and Amazon have done everything in their power — based on discussions with sources familiar with their infrastructure — to obfuscate the actual underlying costs of their operations, and Silicon Valley, an industry of alleged free thinkers and individuals, is more than willing to accept whatever convenient myths might sustain their dreams.  And in the end, they all became useful idiots for hyperscalers. Their obsessive attachment to OpenAI — and by extension Anthropic — seems like a decision made under the auspices of “democratizing powerful AI,” all as effectively every dollar flows to either Microsoft, Google, Amazon, or Oracle, who in turn feed that money to NVIDIA or Broadcom, who in turn feeds that money to TSMC, SK Hynix, Samsung, or Micron.  Invest in an AI startup? They’re gonna be paying one of the AI labs, who will in turn pay a hyperscaler. Invest in an AI infrastructure company? That money will flow to NVIDIA, and then upstream to semiconductor companies. In the end, whether they die or get acquired (as none of them are going public) , all of the value will end up in the hands of one of the hyperscalers who created this imaginary era, then helped inflate it into something very, very dangerous. Yet the problem is that this industry cannot, under any circumstances, survive without OpenAI.  When people discuss OpenAI’s potential collapse, they act with pure cowardice either saying “it won’t be that bad” or say something vague about it “ being too big to fail .”  If OpenAI — the company with the most money and the most infrastructure and the most attention and the most talent in AI — collapses, it will likely do so after AI data center debt and venture capital funding has been almost entirely exhausted.  You see, Goldman Sachs’ Jeffrey Papai recently noted that it will be “very difficult” to replicate the hundreds of billions of dollars that hyperscalers have raised in the last four years — $244 billion in 2026 alone if you include NVIDIA and SpaceX — which is a problem considering that they can no longer fund their data center capex using their cashflows as of Q3 2026 .   And to be clear, hyperscaler capex doesn’t have to stop for NVIDIA to stumble. It just has to slow down meaningfully enough that Jensen Huang can no longer give investors 60%+ year-over-year revenue bumps, because the AI bubble is built on vibes, and it can only survive so long as those vibes don’t become sour.  Yes, yes, I realize there are other customers, but the vast majority of NVIDIA’s demand comes from hyperscalers, who are (for the most part) either building out their operations for OpenAI and Anthropic or simply copying what the other hyperscalers are doing (see: Meta and SpaceX).  Once hyperscalers stop spending money, banks that are afraid of “choking” on data center debt will see that a vast amount of capital is leaving the market and underwrite (or not, as the case may be) deals as such. This will mean, at some point, that both OpenAI and Anthropic will be walking around with their hands out saying “money please!” at precisely the moment that everybody will be cutting back. While NVIDIA might get a little desperate and throw some extra cash their way, if revenues start collapsing, so too will its interest in further inflating the bubble as investors begin to ask whether any of this was real or one large circular financing scam . While this is absolutely a problem for Anthropic — especially after its $35 billion debt deal with Broadcom — it’s much, much worse for OpenAI, which has (as mentioned) made $748 billion in compute commitments to some of the largest and well-lawyered companies in the world. OpenAI’s continued marketing efforts involve constantly refreshing rate limits around the launches of its most-expensive models, giving away millions of dollars of tokens to startups , and generally running the “grow as fast as possible and work out a business later” model into the ground at speed, all fueled and funded by Clammy Sammy Altman’s nasty habit of overpromising and underdelivering. Clamuel’s biggest mistake was leaving the pearly gates of the hyperscalers and dancing with the mortals of Oracle, Cerebras, and CoreWeave. While Microsoft or Amazon might be willing to extend payment terms as a means of saving face and prolonging the inevitable, Oracle — a law firm with a software company attached — is more than capable of loud and aggressive litigation under any contractual breach. Then there’s the fact that Apple is suing OpenAI after poaching multiple engineers for its hardware efforts and allegedly both coaching and coercing them into stealing trade secrets , which is all but certain to destroy any chance of OpenAI releasing a device in the next few years…and potentially the company itself. These are extremely serious allegations, with Apple also accusing OpenAI of trying to coerce trusted partners into revealing manufacturing techniques for iPhones — the kind of thing that can (and will) lead to brutal discovery and potentially criminal charges. OpenAI also, as I’ve mentioned, needs to keep growing to keep up with those bills, and at some point will run out of real dollars to pay people, likely at exactly the time that it’s hardest to find more of them. While there might be billions of dollars left to be raised, to pay any of its bills, OpenAI needs tens of billions of dollars multiple times a year. Based on my own reporting on its audited financials from 2024 and 2025 , OpenAI will need to raise funding at least three more times in the next decade.  To make matters worse, its free users have become a massive liability. While The Information reported that OpenAI expected to generate $2.4 billion in ad revenue in 2026, and $102 billion in 2030 , it turns out that reality is a little harsher, with analyst eMarketer projects that the entire AI chatbot ad industry combined will only make $1 billion this year , with the entire market making $5.41 billion by 2030.  This means that the 900 million weekly active users of ChatGPT will remain a massive drain on the company’s finances, with only 5% or so of them opting to pay , and a projected 80% of its $20-a-month users expected to churn in 2026 . At some point, OpenAI will simply run out of money. It’s nearly exhausted every available source of capital, and now that it’s likely delaying its IPO to 2027 — largely in part because it couldn’t list at a $1 trillion valuation — it will have to raise again, potentially at a down-round valuation or at a modest increase which will, in turn, make it much more difficult for investors to see a return in an IPO.  Investors will likely ask questions like “why couldn’t you go public?” and “what is it that bankers didn’t like?” as Sam Altman looks at them like this: You see, OpenAI is awesome at selling mythology and hype, but crumbles the second that its numbers have to face the cold, harsh light of day.  While it’s been able to skate by in situations like Altman’s ouster and its conversion to a for-profit, these were strictly legal situations that could be dealt with by lawyers and cheered on by the press . OpenAI has never faced a problem like “not being able to pay its bills” or “breaching a contract with a major company,” and I think these are an inevitability in its future. In the end, OpenAI’s collapse will be a dramatic narration of the boring, horrifying economics of the AI bubble. Let me explain: The AI bubble is inflated based on hype and hopium rather than tangible proof or substantial revenues driven to anyone outside of the semiconductor industry, and without NVIDIA’s massive returns, I don’t think anybody would’ve taken it seriously past 2024. Any and all achievements of the AI industry are a direct result of market psychosis, a broken media ecosystem, and a trillion dollars that could’ve been sunk into literally anything else, and must be evaluated as such. The double-edge sword of a mythology-inflated bubble is that it’s much harder to sustain when said mythology dies. The AI bubble was able to grow to such a horrendous size because the markets and the media were willing to accept basically anything that Sam Altman or the greater AI industry said.  By waving away any economic problems as growing pains and dismiss those who would scrutinize it as haters or cynics, reporters and analysts provided investors with the justification to invest again and again in these companies without them ever having to make a real business , which means that, well…they don’t have real businesses, which is a problem when you need to actually pay somebody money that wasn’t given to you by a venture capitalist. This will leave the AI industry short-changed in its most-desperate times.  The media is important for many, many reasons, but one of the biggest ones is that scrutiny is what keeps capital in check, for the benefit of humanity and at times the companies themselves. By choosing to pull their punches, ignore glaring economic problems and accept every projection with blind faith, the media empowers grifting and suffocates good businesses as a result, encouraging bad behavior and helping them raise unbelievable amounts of money at ridiculous valuations without worrying about having to make a good business. In some cases, the media even encourages them to do so, saying that “all startups lose money at first” instead of thinking about things for a fucking second. When companies know they won’t face that scrutiny, they engineer themselves as such, putting off ever finding a real business model in favor of whatever will make them buzzy enough to get coverage and raise funding as a result. In a vacuum of skepticism, bubbles inflate, monsters get rich, and regular people always get left holding the bag. As a result, if companies ever bother to become a real business, they only do so at the very last minute, endangering anyone who has backed them and every counterparty in the event they’re incorrect.  When OpenAI dies, it will be after a prolonged period of desperate reorganization and attempts to appeal to investors and the media that it can, in fact, become a real business. These attempts — price increases, price cuts, selling off IP, nebulous circular deals, and so on — will all fail, and by the end, Sam Altman will have run through every single trick imaginable to keep the party going.  And when those fail, what do you think Perplexity does? How about Harvey? Cursor got the last chopper out of ‘Nam with the SpaceX acquisition (assuming it actually happens), but what, exactly, is Cognition, or Glean, or Sierra, or really any AI startup meant to say to compel investors to believe in them once OpenAI dies? That they’re different? That they’re gonna work it out after the company that got given basically everything it needed failed?  The entire AI industry’s sales pitch is that OpenAI opened the world’s eyes to the power of AI, and that giving the AI industry as much money as possible would end in economic abundance the likes of which we’ve never seen. Instead, we’ve got two AI labs that both lose billions of dollars, and the latest model from one of them randomly deletes people’s stuff. It’s not like any of this was sold on actual ROI or real businesses or returns or productivity or any actual measurable thing other than physical infrastructure erected in its honor.  There are simply no compelling stories about the AI industry that can be told in the present tense. Everything is always based on the theoretical multiplicative power of just waiting a few more years, which becomes much harder to believe if the company with the Mandate of Heaven gets sent to Cocytus.  This will have massive downstream effects on basically everything and everyone connected to the AI industry. You won’t be able to raise money for a startup to spend money on compute, nor will you be able to convince somebody that your LLM wrapper will change the world, nor will you be able to justify a massive valuation. Venture capitalists fancy themselves as brave soldiers of the economy, but are really cowardly lemmings that will sprint for cover the second that things get rough.  I also keep hearing from people that Anthropic is magically safe from the AI bubble’s clutches, or insulated from its rotten economics. The amount of pure mythology and misinformation I read about this company on Twitter is genuinely offensive, and the fact that journalists have categorically failed to push back against it is proof that too few people give a shit about anything other than which boot they get to lick next. Anthropic faces the same economic realities as OpenAI. It burns billions of dollars on training, it hides inference costs in sales and marketing, and the only real differences are that it focused more on coding and made fewer ridiculous infrastructure commitments…right up until this year, when it committed $200 billion in compute and hardware commitments to Google , raised $35 billion in debt from Apollo to buy Google TPUs , signed a $15 billion a year compute deal with SpaceX , and agreed to a 20-year-long, $19 billion lease with TeraWulf . Much like OpenAI, Anthropic is also doing way, way too much. There’s Claude for Life Sciences , Claude for Legal , Claude for Small Business , Claude Design , and even, for whatever reason, reports that Anthropic intends to develop its own drugs — and instead of saying “hey man, what the fuck are you doing?” the media falls over itself to repeat and celebrate every single one as if they’re all viable or useful products. Anthropic is as messy, disorderly and unfocused as OpenAI, but has done a better job of convincing people that it’s somehow “ethical” as it fucks over its partners and farts out 200 new products a month.  This is a company that lacks focus or vision other than “more” and “bigger.” The only thing that differentiates OpenAI from Anthropic at this point is the nebulous promises of “AI code” and Dario Amodei’s Doom Trolling and safety theater. The fact that the majority of the media made no efforts to push back against its shenanigan-rich “profitability” narrative is why we’re in this fucking mess.  Anthropic is an AI lab just like OpenAI. It uses GPUs, TPUs and Trainium chips. It trains models in much the same way to do much the same things, and builds quasi-functional plugins on top of them, just like OpenAI does. It makes big compute commitments, it had its infrastructure built out for it by hyperscalers, its CEO is annoying and beloved by cretins, and its value is largely determined by 1000 people on “X The Everything App” experiencing varying levels of AI psychosis.  Attempts to claim otherwise are tacit admissions that OpenAI is unsustainable. Please note that when I say “victims,” I don’t always mean “people you should feel sorry for.” In some cases I’ll be talking about real people who are facing the horrible consequences of the OpenAI bubble bursting, and for whom you should feel a degree of sympathy, and in others, I’m referring to various Patagonia gargoyles’ financial woes. I assume you’ll be able to differentiate between them.  My last premium newsletter was the massive Hater’s Guide To The Memory Crisis , or the twisted tale of how three companies — Samsung, SK Hynix and Micron — have diverted meaningful amounts of manufacturing supply away from making the RAM you find in laptops and smartphones toward making the high-bandwidth memory that powers GPUs, jacking up the price of consumer electronics in the process.  To explain: To simplify, the AI GPUs in AI data centers require hundreds of gigabytes of high-bandwidth memory, the CPUs attached to them require the same RAM as your smartphone, and the companies making all of this RAM are making huge profits by jacking up the price because of supply chain constraints that they themselves have created. That’s why Micron had 84.9% gross margins in the last quarter . The RAM triopoly controls more than 90% of the world’s memory, and can set prices at whatever rate they want. These three companies were all fined over $100 million by the Department of justice back in 2002 for price-fixing , with Micron avoiding the fine by turning in its co-conspirators . Five years later in 2007, a Supreme Court judgment and resulting precedent ( Bell Atlantic V. Twombly ) drastically raised the bar for not simply winning an antitrust case, but even getting one to trial : This precedent would kill a 2019 class action case against SK Hynix, Samsung and Micron that alleged they had colluded to tighten the supply of the world’s DRAM , because despite statements from company representatives made at public events, their collective participation in certain industry groups, and observable pricing trends, the precedent set by Twombly meant that the plaintiffs required more than circumstantial evidence to bring something to trial.  Anyway, the reason I bring this up is that while I am not accusing Samsung, SK Hynix, and Micron of price-fixing, a recent lawsuit is accusing them of exactly that : So, what does this have to do with OpenAI?  Well, back on October 1, 2025 , OpenAI, Samsung and SK Hynix announced a “strategic partnership” that would involve OpenAI buying 900,000 wafers of DRAM a month (around 40% of the world’s supply at the time) for Stargate data centers — something that never actually happened (it was a memorandum of understanding, and OpenAI also had nowhere to put them), but both SK Hynix and Samsung’s stocks immediately rallied , and Samsung happened to hike prices by 60% a month later , which could be a coincidence, or could have been the company saying “yeah, wow, we’re gonna run out of RAM I guess, better buy now at whatever price we have it!” Another clue that this might not all have been above board was that Samsung was reportedly doing another deal with OpenAI in March 2026 , “...to supply up to 800 ⁠million gigabits (Gb) of 12-layer HBM4 chips to OpenAI in ​the second half of this year” per Reuters, for use with Broadcom’s custom “Jalapeno” chip . Though it’s hard to calculate exactly how much that would be wafer-wise, from what I understand we’re talking in terms of less than 100,000 wafers total after OpenAI, Samsung, and SK Hynix said they’d be taking up 900,000 a month. Regardless of whether OpenAI ever takes a single wafer of silicon, these deals existed to put the squeeze on any company that uses memory in their products — including NVIDIA, AMD and Broadcom — which in turn led to the most aggressive price increases in the history of consumer electronics. As I said last Friday: And yes, OpenAI is responsible, both in its naked collusion with memory manufacturers to push an announcement that never resulted in anything other than price increases and its siren song that made every dimwit with debt desperate to build AI data centers.  Every single consumer suffers as a result. RAM is in everything, and it’s unclear when new manufacturing capacity will actually come online, as fabs are expensive and complex construction efforts and require tons of specialist talent, raw materials, permitting, land and power. SK Hynix Chairman Chey Tae-won said in March that the memory shortage would last until 2030 , and he may be right, as a Bank of America report just said that SK Hynix may only be able to add a sixth of its planned capacity by 2028 . This means that the price of consumer electronics will be inflated for the foreseeable future, even if the AI bubble bursts. While capex pullbacks will eventually happen and by extension eventually lead to supply constraints easing, Micron, Samsung, and SK Hynix had sold out their entire 2026 supply by the second week of January , and noted that they’d only be able to handle 60% of “medium-term” customer memory orders, which suggests to me that 2027 might be even worse, with a subtle clue being that SK Hynix CEO Kwak Noh-jung recently told Reuters that 2027 would be “the worst year in the industry’s history from a supply perspective.”  While the memory triopoly has every incentive to make things seem bleak to drum up business and sustain their margins, behind the scenes reports suggest they’re turning the screws on everybody. This is a graphic example of companies with massive amounts of leverage using it to fuck over both their customers and their customers’ customers .  Who gave them that leverage? The AI industry and Sam fucking Altman.  Hey, remember when I just said that ( it seems, but I cannot confirm that) OpenAI helped SK Hynix and Samsung manufacture a supply chain crisis last year using a phoney announcement for a project that would never happen? That happened three other fucking times in the same three week period, and modern journalism doesn’t seem to give much of a shit! Let’s review what happened, per my year-ending Enshittifinancial Crisis newsletter : All four of these companies’ stocks rallied on deals that land somewhere between misleading and fictional, with basically anyone who invested in them being underwater within two months, though all three have recovered thanks to similarly-questionable announcements and deals made by companies with the sole intention of boosting their stocks.  Why else would Sam Altman go on CNBC with NVIDIA CEO Jensen Huang on the day of an announcement of a project that was only ever a letter of understanding ? Why else would Sam Altman jump on TV with Bob Iger to talk about a Disney deal that clearly never went anywhere? Spare me any explanations around the “fast-paced dealmaking of AI” or “how deals are complex.” CNBC reported the day after the NVIDIA deal was announced that the first $10 billion tranche would “close within a month or once the transaction had finalized” via a source! It’s blatantly obvious that the intention was to create the appearance that a deal existed that never actually existed at all! The AI trade is the natural endpoint of an increasingly-enshittified stock market where many analysts and journalists exist only to repeat narratives to influence stock prices. Outside of semiconductors, the AI trade has never, ever been about the actual underlying economics or the actual economic potential of Large Language Models, but projecting shadows on the wall to resemble something that looks like the next generation of technology. That’s because the AI trade is entirely symbolic and driven by stock prices. When NVIDIA and the rest of the Magnificent Seven (sans Apple) does well, AI is the greatest thing on Earth. When the Magnificent Seven stumbles, everybody worries that they might be overspending on AI. The AI trade exists only to manipulate stock prices through spurious news and smoke signals on social media, and to drag gullible retail investors ( who account for 20% of US equity trading volumes, the highest it’s been since 2021 ) and the rest of the market away from caring about things like “fundamentals” or “reality” toward whatever keys are currently jingling.  My evidence is fairly simple: Google, Meta, Microsoft, and Amazon don’t actually tell you their AI revenues, other than when Microsoft and Amazon have chosen to define it in terms of undefined “run rates.” And why would they? Reporters have been saying that their AI bets have paid off for years without the companies ever having to show it paying off other than their stocks running.  Here’s another example: CoreWeave, a time bomb /AI compute company that only really exists as a revenue source for NVIDIA ( per Jensen Huang , if [NVIDIA] didn’t help CoreWeave exist, they would not exist”) by signing contracts with companies for unbuilt capacity that it then takes to banks and uses to raise more money to buy GPUs. NVIDIA knows that analysts and reporters don’t give a shit about the blatant self-dealing and circular financing, all because these deals help the stock price go up, which apparently is the only metric that modern journalism evaluates. That’s why when NVIDIA invested $2 billion in CoreWeave in January 2026 — a warning sign that the company had liquidity problems! — led to endless positive coverage after “the stock popped on the news,” per CNBC. That’s because the AI trade exists only to extract value and con investors. It is not a trade related to the actual fundamentals of whether AI works or not, whether AI actually makes anyone money, or really anything about AI at all outside of whether mentioning AI or an AI-related company makes a stock number go up or down. I’ll be blunt: modern journalism has failed the retail investor and directly helped the wallet inspector regulate the stock market. By empowering Sam Altman and the rest of the AI industry’s deliberate attempts to obfuscate the actual economics of generative AI and setting the terms of AI’s success as “how stocks are doing and whether the companies are growing in general,” they have defaulted on their responsibility to the general public and helped the already-rich get richer.  None of this would be possible if business journalism actually saw themselves as having a responsibility to give their audience good information. While one could argue that if you had blindly invested in the AI trade you might have made money, the ability to make money in the AI trade was directly driven by modern journalism’s inability or unwillingness to push back on any corporate narrative. Every major outlet ran a story on every one of the deals I mentioned, and not a single one seemed remotely upset or deterred by the fact they were misled, and in turn misled their audience. And yes, investment funds can be just as easily manipulated as a retail investor, and will follow whatever trend seems likely to make them money, even if said trend is utterly disconnected from any fundamentals. Tech analysts help do so by creating vast models that give a veneer of respectability, even if their projections mostly amount to “number will always go up in the future.”  This is why Musk was able to dump SpaceX on the public markets. Why SK Hynix chose to list on the NASDAQ. When the entire world is captured by a childlike belief that “AI is good and will be the biggest thing ever,” you empower grifting and swindling at scale.  Well, that and underwriters like Goldman Sachs are so nakedly crooked that they’ll say they expect SpaceX’s AI revenue to grow 100x by 2030 . Fuck off! Yet the memory boom/bust/crisis is where the media has failed investors the most — a final insult before everything collapses. You see ( to quote myself ), what makes this particular memory crisis so distinctly dangerous is that it isn’t a result of consumer demand so much as it is capital expenditures from very large companies making bets that don’t connect with reality.  Microsoft, Google, Amazon, and Meta aren’t spending $765 billion in capex in 2026 because of rapid demand by consumers for AI services, but a desperation caused by a lack of hypergrowth ideas , circular financing with Anthropic and OpenAI , and a vague concern that if they stop spending that the other guy will do something as a result.  Anyone blathering on about a “memory supercycle” is intentionally obfuscating where that revenue and demand is coming from — high-bandwidth memory attached to AI GPUs, meaning that this boom cycle only exists as a symptom of a greater hype cycle, meaning that when companies stop buying GPUs , the demand for that (briefly) high-margin high-bandwidth memory goes with it.   To give you some context, a chart from ComputerBase.de showed that high-bandwidth memory demand grew from 681 million gigabits of HBM in 2022 to 29.3 billion gigabits on 2026 — a 40x increase over the course of four years that suggests that once GPU-related capital expenditures stop, high-bandwidth memory demand will effectively disappear .  As I mentioned previously, this isn’t even me being a hater . Hyperscalers are now joining the rest of the world in having to raise debt to buy more GPUs, which means that at some point they aren’t going to be able to afford to buy as much, which will in turn mean that NVIDIA — which accounts for around 65% of all HBM purchasing — won’t need as much. I have not read a single fucking article that mentions that this is a possibility! Every article about the memory industry right now is about supply constraints and the increasing cost of memory , but none of them warn investors or the general public about what will happen when capex slows , and certainly not the many, many articles in major business publications about SK Hynix, Samsung and Micron’s revenues. In fact, Reuters said that SK Hynix’s “ scarcity premium looks built to last .” The cynical (and boring) response here is that “the market can stay irrational longer than you can stay solvent,” but saying that distracts from the larger point of how said irrationality was manufactured by the media .  I am not sure what the majority of the media sees as its purpose or responsibility to its readers, so I will speak plainly: the responsibility is to tell them the cold, hard truth, rather than going along with whatever hype cycle is happening out of fear of being wrong or missing out. Skepticism is not doomerism! Being critical is not being negative! These companies are some of the largest and richest enterprises in the world — they should be scrutinized! And no, scrutiny is not publishing everything they say and then making a vague comment about “whether or not that bet will pay off.” Too often, journalism conflates objectivity with passivity, seeing critiques as “negative” or “biased” when, in fact, repeating everything that corporations say to their benefits is about as biased as it gets. In the end, the victims are anybody who doesn’t exit the AI trade in time.  By the way, there’s no Hell hot enough, by the way, for the people that will read this and smugly say “heh, well, I made money,” or who point to anyone’s returns as evidence that the AI trade is anything other than manufactured consent. The fact that anyone made money on this trade is a sign that the stock market is inherently manipulated to benefit the wealthy at the cost of the many — and when the bubble bursts, the people that will suffer will have suffered because of the media’s participation by helping Sam Altman and the rest of the AI industry obfuscate and twist reality to pump stocks. Which leads us neatly to our next victim! In my Hater’s Guide To SoftBank , I told the story of CEO Masayoshi Son, a degenerate gambler who has steered his company through boom and bust cycles only through the grace of whatever God he believes in and sheer luck.  SoftBank Group — the holding company, and not to be confused with Softbank Corp, which runs a bunch of telcos and media companies in Japan — makes money only through either investing in or buying companies, then taking them public or selling them to someone else, and otherwise needs debt for liquidity.  Masayoshi Son makes terrible bet after terrible bet, but his luck always seems to work out for him. His $20 million stake in Alibaba turned into $50 billion at IPO. He bought a 70% stake in Sprint that turned into a 24% holding in T-Mobile . In the early 2000s, Softbank took a 23% stake in Betfair that eventually became part of the $17.7 billion Flutter Entertainment. And then there’s its most-recent and arguably most-impressive (after Alibaba at least) investment, ARM, which it acquired for $32 billion in 2016 and then took it public in 2023 at a valuation of $54.5 billion , and currently sits at around a $300 billion market cap.  Yet his problem has always been his dalliances with whimsical white boys. SoftBank sunk $1.5 billion into dodgy financial services firm Greensill Capital before its collapse, and in the aftermath, it was revealed that Masayoshi Son and CEO Lex Greensill talked on the phone every day , to the point that ( per Greensill himself ) SoftBank managers felt “threatened” by Greensill’s relationship with Son. It only took Masayoshi Son 28 minutes of conversation with WeWork’s Adam Neumann before he drew up the terms for a $4.4 billion investment on his iPad and signing the deal in the back of a cab, with Son saying that “the last person he felt this with was [Alibaba CEO] Jack Ma.”  And no white boy has ever been more whimsical than Sam Altman.  In 2019 , Altman turned down $10 billion from Masayoshi Son (which, ironically, would’ve been an incredible investment at the time), going instead with $1 billion (and full infrastructure support) from Microsoft, and I believe this moment drove Son into a level of madness that will potentially wreck the company. You see, up until fairly recently, SoftBank had been dragged down by the declining value of its atrocious investments via its two venture capital funds — Vision Fund 1 and 2, the latter of which was self-funded and has mostly gone toward funding OpenAI. Up until recently, SoftBank had quarter after quarter of losses as investment after investment saw its NAV drop because, well, they were overvalued and SoftBank never should’ve invested in them in the first place. To survive, SoftBank moved into “ defense mode ” in 2020, slowing investments and selling the vast majority of its Alibaba stock by April 2023 , with the ARM IPO and billions of dollars of bond sales helping slow the bleed. Yet Masayoshi Son knew he was destined for greater things, as he told CNBC in June 2024 : OpenAI — and the larger AI trade — had given Masayoshi Son a certain kind of greed-driven mania, where he believed that AI would make SoftBank (as he said recently) “ the goose that laid golden eggs ,” an eternal money-printer that ostensibly started with the biggest cash-burning machine in history.  Altman, like Neumann, like Greensill, told Masayoshi Son exactly what he wanted to hear: that this would be the biggest thing ever, and that Son would capture all of the value both through his investment in OpenAI and further investments in data centers and other AI infrastructure.  And so began his most vulgar investment yet — OpenAI, sinking $2 billion into the company from Vision Fund 2 in November 2024 — only for Altman to turn around and demand he fund $30 billion of a $40 billion round that would get announced four months later in March 2025 . Masayoshi Son was an emphatic “yes,” except for one little problem: he didn’t have the money, and could only afford the first $7.5 billion (due in April 2025) by taking out a $15 billion, year-long bridge loan , with the rest of it going toward his eventual purchase of Ampere computing .  To fund the remaining $22.5 billion, SoftBank was forced to take out further margin loans on its ARM stock , and sell large chunks of its T-Mobile stock , as well as its entire $5.83 billion stake in NVIDIA . Yet as soon as the check cleared, Sam Altman was blowing up his phone demanding more money as part of a $110 billion funding round in February 2026 (that eventually became $122 billion in late March). Masayoshi Son was once again an emphatic yes, except by this point he’d exhausted basically every useful thing left in his coffers outside of around $118 billion in ARM shares that make up around 40% of SoftBank’s net asset value, meaning that selling or using further ARM shares as collateral would directly tank its value — both through the obvious “they have less of a valuable thing” and sales/collateralization of further ARM shares affecting its share price. So, what did Masayoshi Son do? More debt, baby! More risky debt! You can always refinance it, right?  To pay for its share of OpenAI’s 2026 funding round, SoftBank took out a $40 billion bridge loan (maturing in March 2027), bringing its investment in the company to over $40 billion, with its payments to $10 billion tranches of OpenAI funding due in April, July and October 2026. A few months later, it tried to raise a $10 billion margin loan using its entire OpenAI investment as collateral, cut the amount it was raising to $6 billion, and when banks remained hesitant to give it the money anyway offered to “ guarantee repayment of the loan to address lender concerns, ” effectively backing the loan with its own balance sheet (called a recourse loan) because, despite being worth over $100 billion on paper, its lenders had doubts that its OpenAI stake was actually worth that much.  If you’re wondering why it didn’t simply take out more debt, it’s because (as a result of its continuing investments in OpenAI) S&P Global revised SoftBank’s outlook to negative , emphasis theirs : This has had a knock-on effect on the rating of the telecoms-focused Softbank Corp (as a reminder, Softbank Group is the holding company that owns stock in other companies, Softbank Corp is the energy/telecoms company that actually makes stuff), which is now rated BBB, or the lowest-possible rung of investment-grade financing in the S&P system. To make matters worse, if SoftBank continues to hold a loan-to-value ratio of above 30% for much longer, it runs the risk of its debt getting downgraded even further, which would slam the door shut on its ability to raise money via bonds, which is…well, basically how SoftBank has functioned for the last 10 or 20 years. And this is all happening as Japan is determinedly inching away from the era of persistently low interest rates — making debt far more expensive to service.   SoftBank needs OpenAI to IPO so that it can turn that on-paper gain into actual liquid stocks that can be dumped into the market or used for real-life margin loans. SoftBank has jettisoned the vast majority of its heaviest-weight investments, leaving it largely dependent on the continued value of ARM’s stock to keep its seat at the table, and if OpenAI can’t go public, it’ll end up sitting on illiquid stock in a company that will see its value tank as a result. Yet even if OpenAI does go public, any attempts to get a margin loan will likely be dangerous, as I bet that it will be one of the single-most shorted and volatile stocks in history, which will also be a problem for SoftBank’s underlying net-asset value, which will ebb and flow based on whatever bullshit Altman cooks up every three months. Masayoshi Son is both a victim of the manufactured consent of the AI trade and an enabler of its worst excesses, empowering and enriching Sam Altman at a time when any kind of financial prudence might have curbed OpenAI’s greed or killed it before it caused further damage.  SoftBank tanking will fuck over anyone invested in the Japanese stock market, where it currently sits as the third-largest company by market cap behind KIOXIA (a memory company booming thanks to the AI trade) and Mitsubishi UFJ Financial (a bank with heavy ties to the AI industry and data center infrastructure). While I severely doubt it’ll die — it’s likely MUFJ and SMBC Bank would extend whatever credit necessary to keep the doors open — OpenAI and the greater AI trade has become a load-bearing toothpick holding up the trillion-ton ass of the world’s most well-funded gambler. For SoftBank to survive in its current form, OpenAI must go public, become a thriving and profitable business, and have its stock price stay elevated for the foreseeable future. Additionally, ARM must also retain or exceed its current stock price. Hey, while we’re on the subject of “companies betting the entire future on OpenAI that recently got downgraded by S&P Global…” Hey! You in the back! Stop laughing! Stop laughing at Larry Ellison! He’s now only the world’s 8th-most-richest guy !  Just kidding, fuck Larry Ellison. What I’m about to tell you might make you laugh, probably because it’s really funny. Oracle is currently spending over $340 billion to build out over 7.1GW of data center capacity for OpenAI , as part of its $300 billion, five-year-long cloud compute contract that began, at least in theory, on June 1, 2026 at the beginning of its Fiscal Year 2027, though much of the capacity is yet to be built. To fund the buildout, Oracle has had to raise over $50 billion via stock sales and debt , spent $55.7 billion in its last fiscal year , and expects to spend at least $90 billion more in FY2027. As a result of that , S&P Global downgraded Oracle’s credit rating to BBB/A-2 , the literal lowest level before it’ll become junk-grade, meaning that one more downgrade ( though it would have to be from two ratings agencies ) from here would risk Oracle becoming a “fallen angel,” with investment funds (that can’t hold junk grade debt) having to jettison its debt from indexes, as happened to Ford in March 2020 , leading to over $35 billion in debt being dumped and its borrowing costs skyrocketing to between 8.5% and 9.625% when it raised in April 2020 . For some context, Ford reported an average interest rate of 5.2% on its long term debt in its 2019 annual report .    You’ll never guess why S&P Global downgraded Oracle! And, once again, the emphasis is theirs: That’s a load-bearing if, brother!  Anyway, you know who else is trying to warn you about Oracle’s exposure to OpenAI? Oracle! Per Bloomberg : As a reminder, the only way that OpenAI will be able to afford to pay its $300 billion cloud compute contract with Oracle will be if it continues to hit revenue projections ( per The Information ) that have it making $113 billion in 2028, $184 billion in 2029, and $284 billion in 2030, a year when it will magically become profitable, and no, I don’t know how that happens: Based on my own analysis , assuming that Oracle can successfully build capacity for OpenAI to pay for (a load-bearing assumption), it would have to pay around $75 billion to rent that 7.1GW of capacity. Stargate Abilene, an 8-building, 1.2GW project that broke ground in July 2024 , has (per sources familiar with the matter) only built and operationalized three buildings, despite the project having meant to be fully operational by the end of 2025 ( per landowner Lancium ), or energized by the middle of 2026 , it isn’t really clear, and I can’t get a straight answer from anyone about whether the power even exists on site to turn any of it on.  Anyway, for Oracle to make all the rest of that money, it will have to build five more Stargate Abilenes. If you’re wondering how that’s going, Stargate Shackelford only broke ground in December 2025 , Stargate Wisconsin appeared to have a single steam beam in March , Stargate Michigan only got its first steel beams two months ago , and Stargate New Mexico is still waiting for permitting to begin construction .  Based on Lancium’s presentation and discussions with sources familiar, Oracle will pull in somewhere in the region of $10 billion in annual revenue from the (assuming it’s ever done), completely-finished 824MW of critical IT infrastructure at Stargate Abilene. It is unclear how Oracle hopes to be paid even a fraction of its $300 billion compute deal, because in its current state, its annual revenue from Stargate projects currently sits in the region of a maximum $5 billion a year, or less than a tenth of its FY2026 capex. For the most part, Oracle has funded the various Stargate data centers with project financing, meaning that a nebulous SPV will be responsible in the event it defaults on any of these contracts…until Stargate Michigan, which only closed when Oracle agreed to guarantee the $14 billion in bonds raised .  All of this revenue — both theoretical and otherwise — sits in Oracle’s “Cloud” segment, the only part of the business that’s actually growing , as the rest of its business has either been declining or plateauing for about a decade.  In any case, for Oracle to actually get paid its $300 billion, it will have to build upwards of 6GW of data center capacity…in a year and a half? This deal is meant to be worth in the higher range of tens of billions of dollars in annual revenue by FY2028, which begins on June 1 2027! Stargate is horribly, impossibly delayed, to a level that makes me wonder if anybody other than perhaps Anissa Gardizy has bothered to think about Stargate for even a fucking second. Anyway, Oracle’s entire future rides on this deal. While Oracle Cloud Infrastructure continues to grow, its future growth (and remaining performance obligations) almost entirely hinge on both its ability to build the largest infrastructure project of all time and for OpenAI to continue raising funding for an indefinite amount of time. The rest of that growth comes from Meta and xAI, both of whom are only really “doing AI” because everybody else is. This puts Oracle in a very, very compromising position on multiple different levels.  Generative AI is the only reason that Wall Street started liking Oracle again as its other business plateaued, even as it burned billions of dollars on capital expenditures and cut its gross margins by a little under 15% since 2022 , with the vast majority of that value coming from its revenue from OpenAI and what’s actually active at Stargate Abilene.  Much like the rest of the AI trade, everything about Oracle’s future is sold on potential rather than anybody thinking about reality or things like “whether Oracle can actually build the data centers” or “how Oracle makes any of that revenue if the data centers aren’t built” or “how OpenAI affords to pay for the compute if the data centers get built.” As Oracle said in its own disclosures, if OpenAI can’t pay, “Oracle could be left with massive data center leases that it might be unable to exit or have to re-lease to new tenants under less-favorable terms,” and there isn’t a single company on Earth who can or would pay for such a large amount of compute, nor is there the aggregate demand to justify it. While its many government contracts and national security significance make it unlikely that Oracle would be allowed to die , the collapse of its only growth segment will likely spell dark times for a company that’s already laid off 21,000 people as a means of funding its AI buildout. The double-edged sword of the AI trade’s childlike attachment to stock valuations poses an egregious threat to Larry Ellison himself. Hey — HEY! I said no laughing! Stop it! This is all very serious! This is a serious situation! You’re laughing about the potential downfall of a guy who once wrote a letter to the New York Times attacking HP for firing former CEO Mark Hurd for repeatedly making sexual advances toward a reality star using HP’s finances !  Sorry, my mistake, you should keep laughing, even the prospect of what I’m about to tell you is hilarious.  As I said in my piece about how OpenAI Kills Oracle:  One of the consistent themes of this piece is that much of the “value” of AI is hot air — by which I mean whatever people are willing to pay for a stock that’s continually inflated by specious media-driven hype.  Ellison’s wealth is driven by both his share of Oracle’s ongoing yearly dividend, his Oracle shares, and his ability to offer said shares as margin loans, which makes him vulnerable to even a symbolic collapse of OpenAI, which is why it had to tweet in February that “ the NVIDIA-OpenAI deal has zero impact on its financial relationship with OpenAI ” to calm those dumping the stock.  To be clear, Ellison has around 1.16 billion Oracle shares, leaving him with around 810 million or so left, allowing him to pledge them as further collateral rather than having to either dump them on the market or dip into his reserves of about $10 billion in cash and $15 billion in Tesla stock , with Ellison historically never selling more than about $4.7 billion in stock. We don’t know the exact scale of terms of his personal loans, but do know that he’s got a shit-ton of them, and that his entire fortune rests on the idea that he never has to sell Oracle stock. That becomes a problem if things drag on with the Warner Bros deal, as he’s also guaranteed $40 billion from the Ellison Trust , effectively barring him from selling or using those shares until the deal clears (and the money from the Middle East arrives to fund the deal). The amount of shares that Ellison has committed has oscillated on a year-by-year basis, sitting at 305 million in both 2018 and 2019 , rising to 317 million in both 2020 and 2021 , dropping to its lowest level in 2024 ( 217 million ) before bumping back up to 346 million in 2025. While the board theoretically keeps an eye on his loans and what he’s pledging, he holds 40% of Oracle’s stock and the undying loyalty of veterans like former CEO Safra Catz and co-CEOs Clay Magouyrk and Mike Sicilia. To get specific about how the Paramount/Warner Bros deal breaks down, $24 billion will be covered by funds from the Middle East (primarily sovereign wealth funds), with Ellison providing $22 billion and bank debt funding the rest.  If the deal doesn’t close by September 30, the Ellisons have to pay around $650 million a quarter in fees . If it does , Ellison will likely either have to liquidate his Tesla stock, hand over cash, or take out further margin loans on his Oracle stock to fund it. Those would likely increase the amount of shares he’d have to commit somewhere between 150 million and 300 million (at a loan-to-value of 25% to 50%) at whatever price Oracle is currently trading at. Though it’s hard to tell exactly, the number to look for with Oracle is “below $70.” Once that happens, Ellison will likely have to proffer more Oracle stock to keep up with his margin calls, which will severely limit his ability to take out further margin loans using his Oracle stock. He will have to renegotiate loans, and if he’s managed to buy Paramount, he’ll be sitting on the stock of a company with $80 billion in debt and constantly loses money , which will be far less-appetizing to potential lenders who are aware that the rest of Ellison’s money is tied up in the plummeting hopes of Oracle. Things could get much darker if Oracle plunges below $50, as at that point the encumbrances of his various enterprises and his own margin loans could become too much to avoid having to liquidate Oracle stock. If that happens, it creates a vicious cycle that will potentially involve selling off Paramount, dumping further Oracle shares, or even trying to engineer a firesale for the company. All of this was entirely avoidable if he had never met Sam Altman, and never gave in to the temptation of the AI trade. When the OpenAI Bubble — and OpenAI itself — bursts, many will attempt to eulogize the situation in terms of how we could’ve possibly known this would happen, and I want to be clear that I’m going to be reading and commenting on as many of them as I can find. I believe that once OpenAI collapses it’ll have a violent, punishing effect on the entire stock market, a precursor to a much greater drawdown as everybody accepts that the AI bubble has burst.  This view is shared by the Bank of England governor Andrew Bailey , who warned that the bursting of the AI bubble would have an effect on the UK economy, even though the UK economy — and the UK financial system — isn’t nearly as exposed to it as that of the United States, and would have significant enough effects to change British monetary policy, specifically, interest rates.  And I continue to stand by my belief that this company will die, though I can’t say when it’ll happen. The promises that Sam Altman has made at the scale that he’s made them are equal parts ridiculous and dangerous, leaving any counterparty somewhere between burned or destitute as a result.  There is no compelling story for any AI company once OpenAI dies. Other AI labs will suddenly have to explain how they avoid the same economical pitfalls while still showing the same aggressive growth projections promised by Sam Altman, and half-measures will no longer be acceptable. Their ability to secure credit — or even venture funding — will be met with impossible-to-answer questions about sustainability and profitability. Any startup connected to its models will suffer because it’ll be clear that any AI lab is a financial black hole, and it’ll become obvious that basically every AI startup is an unprofitable LLM wrapper. That should be obvious now, but nobody bothers to look. Any AI infrastructure company will have to pivot aggressively to open source models if they haven’t already, and realize that much of the demand for AI services came from brainless curiosity driven by the AI trade and market hype. CoreWeave, IREN, and the many circular-financed neoclouds will, much like AI labs, find themselves unable to secure funding, as the first question will be “how do you know your customers won’t die?” NVIDIA just won’t be able to justify selling as many GPUs, as it has repeatedly cited OpenAI (albeit without saying its name) as a proxy driver of sales via counterparties including Microsoft and Amazon. It’ll be a permanent blemish on a startup ecosystem that helped so many people become rich based on fictional or fanciful promises and projections, enabled and funded by venture capitalists that didn’t force founders to make stable or sustainable companies because it “always worked out before.” And I genuinely think this will create an accountability crisis in the media. I speak with readers and listeners every single day that are horrified about how many half-truths and outright lies are published and used as a means of propping up the AI bubble and the larger tech industry. The term “AI” has grown from a kind of technology to a cudgel wielded by the powerful to threaten and terrorize workers, all based on the outcomes from Large Language Models that simply do not do what their progenitors have promised and do not produce ROI or productivity benefits that are in any way measurable. The OpenAI Bubble inflated not because Sam Altman is a super-genius, but because he’s very, very good at telling people what they want to hear. He’ll give members of the media convincing-enough projections, said with the confidence ( or necessary fear ) necessary to sway the vast amounts of reporters who are excited to follow the next big hype cycle (or, put another way, are scared to miss out on it).  Altman knows the exact signifiers to use and the minimum viable product necessary to “prove” OpenAI’s worth — however many hundreds of millions of weekly active users, annualized run rates, gigawatts of data centers, vague promises of “abundance” and “intelligence too cheap to meter” that never actually resemble a tangible thing — that work to con reporters and investors who don’t want to think about anything but growth.  He’s also really, really good at playing on people’s greed, be it promising Satya Nadella he can build the next generation of cloud compute cash, Larry Ellison that he can make OCI bigger than Azure, and Masayoshi Son that he can birth a goose that lays golden, AI-labeled eggs.  Altman realized early on that the only way to sell AI was to talk about it in the future tense in a mixture of threats and promises, always subtly suggesting that those who follow the OpenAI gospel will be saved from the permanent underclass. And that same con worked on the minds of Silicon Valley founders who feel sore that they’ve yet to become an early employee of the next Apple, Google, Amazon or Microsoft, selling the dream of endless wealth under the auspices of “accelerationism” that really means “growth at all costs, usually billed to somebody else.” He and his acolytes have created a palpable mania in the Valley, convincing people that not using his software is a guarantee that they’ll be poverty-stricken imbeciles, and I think he’s fully aware of the fact that Silicon Valley is a dense monoculture that LARPs as a free thinker’s paradise. In the end, Altman is unlikely to suffer, at least anywhere near as much as those he’s misled or helped mislead. The scale of losses that the stock market may face scare me to the point I’m almost hoping I’m wrong, with the markets heavily dependent on eternal growth of the AI trade, as without NVIDIA selling more GPUs every quarter, it’s unlikely that anybody is going to be excited to invest in tech past the year 2028. All of this could’ve been stopped if those responsible for scrutinizing the powerful actually did their jobs, and spent more time doing that than critiquing the critics and repeating the promises of craven liars and billionaire scumbags. There were signs from the earliest days that this was all unsustainable, and the only reason it got this big was because the media and the markets fell behind a specious AI trade, empowering and enabling venture capitalists and hyperscalers to sink hundreds of billions of dollars into a doomed industry.  Whatever the AI industry achieves by the end of this farce will pale in comparison to the massive harms it has caused and will cause as a result, and for us to avoid this happening again, we need a fundamental reimagining of how the powerful are covered, how much effort is made to pry apart their plans, and accountability for those who either failed to stop them or actively assisted them. If I sound salty, it’s because I am worried about the regular people caught up in this madness — the tens of thousands of people that have suffered AI-washed layoffs , the hundreds of millions of people that invest in a global stock market dependent on the AI trade (for a taster, see what’s happened in Korea when the KOSPI dropped earlier in the week, forcing hundreds of thousands of retail investors to face margin calls ), those whose retirements and pensions and insurance annuities are tied up in private credit funds invested in AI data centers , and anyone of any kind who built their life around any promise made by Sam Altman and those that followed him.  I challenge those who are glibly dismissive of everything I say — who look for any smidgen of proof to dismiss hard numbers or clear economic issues — to truly think about the consequences of what I’ve written, and take the risk of the OpenAI Bubble seriously. Tech companies are not your friends, venture capitalists are not your saviors, Sam Altman doesn’t care if you live or die, and the AI industry — and Silicon Valley — will dump you the second that you stop being useful as an acolyte or booster.  I love technology, and credit it with making me a success and the person I’ve become, as well as connecting me to many people I love dearly. I believe that tech should be something that empowers, protects and enriches the human experience, something that’s sustainable and reliable and replicable and stable and makes human beings the same as a result.  The tech industry as it stands shows nothing but contempt for the user. Every tech product is somewhere between broken and buggy. The people that write about tech write for the companies far more than they write for those that pay them. Venture capitalists fund companies that they think they can sell to other companies or take public, which in turn means they fund things that are only attractive to people on Twitter or other venture capitalists. Big tech is unregulated, unrestrained, and works entirely to either enrich or fuck over shareholders depending on the day, and because the finance media has little interest in pushing back, they’ll continue to do so to the detriment of the markets and the retail investor. Everything comes back to a distinct selfishness and lack of responsibility across basically every part of the tech industry. The fact that AI has grown this large is a symptom that Silicon Valley needs to be restrained — that it can and will release dangerous, unreliable, unpredictable and unstable products at scale with little regard for the consequences, in part because it knows the media will celebrate it doing so if it can show user or revenue growth.  OpenAI is the company the tech industry deserves — a directionless company of questionable worth that grew in a vacuum of responsibility that exploits greed and ignorance at scale. And the tech industry will deserve exactly what it gets for coddling Sam Altman, and letting his empire grow this large. If you liked this piece, you should subscribe to my premium newsletter. It’s $70 a year, or $7 a month, and in return you get a weekly newsletter that’s usually anywhere from 10,000 to 18,000 words, including vast, detailed analyses of the biggest events and companies in the AI bubble.  As a reminder, if you sign up between now and XX July, you’ll get $10 off a subscription.  OpenAI’s collapse will be a direct result of its loss-laden economics — its doomed, loss-making subscriptions, its pathetic advertising revenue, and API costs that became a “huge issue” for its enterprise customers — and the fact that outside of the hype , AI lacks measurable ROI . When OpenAI eventually leaves CoreWeave, Cerebras, and Oracle in the lurch, there won’t be anyone else to pick up that compute.These are all debt-laden companies, and without meaningful revenues, they’ll struggle to service their obligations. When OpenAI dies — likely folding into Microsoft in the process — it will massively pull back on any and all compute demands, with the likely end of and free ChatGPT and a massive price bump across the board. OpenAI’s demise would also naturally call into question the rationality of investing in any AI startup. If the largest, best-funded, best-resourced company in the entire industry backed by the world’s largest software companies couldn’t make it, why would you believe somebody else would do so? The collapse of the largest company in the ecosystem would also seize up any and all AI data center debt (if any exists at that point), because the literal largest consumer of AI compute would be dead. On September 22, 2025, NVIDIA announced a “strategic partnership” to invest “up to $100 billion” and build 10GW of data centers with OpenAI, with the first gigawatt to be deployed in the second half of 2026. Where would the data centers go? How would OpenAI afford to build them? How would OpenAI build a gigawatt in less than a year? Don’t ask questions, pig!  NVIDIA’s stock bumped from from $175.30 to $181 in the space of a day. The media wrote about the story as if the deal was done, with CNBC claiming that “the initial $10 billion tranche [was] expected to close within a month or so once the transaction has been finalized.” I read at least ten stories that said that “NVIDIA had invested $100 billion.” This deal never happened. Three months later, the Wall Street Journal said that it was “on ice,” and two months after that , NVIDIA pledged to invest $30 billion in the company , and though NVIDIA mentioned investing $18.6 billion in “private companies and infrastructure funds…[including] AI model makers that may indirectly purchase or use our products in the cloud,” it’s unclear how much made it to OpenAI. On October 5, 2025, AMD announced that it had entered a “multi-year, multi-generation agreement” with OpenAI to build 6 GW of data centers, with “the first 1GW deployment set to begin in the second half of 2026,” calling the agreement “definitive” with terms that allowed OpenAI to buy up to 10% of AMD’s stock, vesting over “specific milestones” that started with the first gigawatt of data center development. Said data centers would also use AMD’s yet-to-be-released MI450 GPUs. The deal would, per Reuters , bring in “tens of billions of dollars of revenue.” AMD’s shares surged by 34% , with analyst Dan Ives of Wedbush saying that this was a “major valuation moment” for AMD.  I can find no tangible evidence that OpenAI has bought a single AMD GPU. While its most-recent 10K references a “product purchase agreement with OpenAI OpCo LLC,” and while you can sort of blame the rumoured delays of the MI450 GPUs OpenAI is supposedly buying , it’s weird that AMD hasn’t loudly mentioned this on every earnings call. It’s also weird that in February 2026, Meta and AMD signed a near-identical agreement . On October 13, 2025, Broadcom announced a 10 gigawatt deal with OpenAI , claiming that it would deploy 10GW of OpenAI-designed chips, with the first racks to deploy the second half of 2026 and the entire deployment completed by end of 2029. Broadcom's stock popped by 9% on the news about the 10GW deal, with CNBC adding that " the companies have been working together for 18 months . [emphasis mine, for a reason that will soon become obvious]"  On May 7, 2026 , The Information reported that Broadcom and OpenAI had yet to work out how to finance the initial purchase of its specialist chips. On June 24 2026 , OpenAI and Broadcom would announce the chip had been “developed from design to production in nine months,” the kind of blatant lie that you tell when you know nobody in the media is watching.  On December 11, 2025, The Walt Disney Company announced that it had reached a “ landmark agreement ” with OpenAI to bring its characters to Sora, adding that it would invest $1 billion in the company. The same day, Disney CEO Bob Iger and Sam Altman went on CNBC , with Iger adding that Disney “[wanted] to participate in what Sam is creating, what his team is creating,” and added that Disney “thought this is a good investment for the company.” It would also buy ChatGPT for the entire company. On March 24, 2026 , OpenAI announced Sora was dead, the deal was dead, and it’s unclear whether anything actually happened.

Finance AI Dart

▲ 0

0 views

Max Bernstein 1 months ago

A survey of inlining heuristics

Compilers, especially method just-in-time compilers, operate on one function at a time. It is a natural code unit size, especially for a dynamic language JIT: at a given point in time, what more information can you gather about other parts of a running, changing system? I don’t have any data to back this up—maybe I should go gather some—but on average, methods are small. Especially in languages such as Ruby that use method dispatch for everything, even instance variable (attribute, field, …) lookups, they are small . And everywhere. This makes the compiler sad. If we are to continue to anthropomorphize them, compilers like having more context so they can optimize better. Consider the following silly-looking example that is actually representative of a surprising amount of real-world code: Right now, in the method, I count 8 different method calls: (Technically more, but the ivar lookups (including !), addition, and subtraction are generally specialized and don’t push a frame, even in the interpreter.) Furthermore, there are at least two heap allocations: one for each instance. Last, there is a bunch of memory traffic to and from instances. This all is a huge bummer! What should be a simple math operation is now overwhelmed with a bunch of other stuff. is certainly not a zero-cost abstraction. Even if we had a bunch of other optimizations such as load-store elimination or escape analysis, they would not be able to do much: pretty much everything escapes and is effectful. That is, unless we inline . Inlining is the lever that enables a bunch of other optimization passes to kick in. I wrote about the design and implementation of Cinder’s inliner ( FB link , personal blog link ) a couple of years ago. I wrote about arguably the simplest part, which is copying the callee body into the caller. It took me at least a week to get working. Probably closer to months if you consider all the plumbing through the rest of the JIT. In February during a small hackathon, I watched my colleague k0kubun prototype that bit of the inliner inside ZJIT in about 30 minutes. There is more to do when pretty much every part of the VM is observable from the guest language: both Python and Ruby allow inspecting the state of the locals, the call stack, etc from user code. Sampling profilers also expect some amount of breadcrumbs to work with to inspect the stack. So there’s some more machinery still required to pretend like the callee function was not inlined. I talk about this a little bit in the Cinder blog post. Even so, all of that can probably be designed and wired together in a couple of months. Then you will find yourself tuning the inliner for the next 10 years. This is much harder. The thing that makes inlining difficult, especially in a method JIT, is that you are trying to make an entire (dynamic!) system faster but you are only looking through a microscope and only capable of local reasoning 1 . Whereas other optimizations such as strength reduction, inline caches, and value numbering are an un-alloyed good for the generated code, inlining can have negative effects . It is also perhaps the first optimization people add that has non-local impact. If you inline wrong, your code size might blow up. This might thrash your CPU’s caches. Bummer, but happens to the best of us. But also, if you inline wrong, you might get in the way of other helpful optimizations: if you hit some size limit after inlining method A, you might never get to inline B, which is the key to unlocking the performance of the method you are trying to optimize. Last, inlining might hurt compile time. In situations where latency is paramount (think: interactive client JavaScript), adding tons more code into the fray might add noticeable hiccups, even if the long-term throughput improves. As always, in-band compilation is a trade-off because any time you spend compiling, you are not executing code . You have to write your compiler to reason about all of this stuff. So you have heuristics. For example, here is Michael Pollan’s inliner heuristic: Inline methods. Mostly small. Not too many. I did a survey of a bunch of compilers, mostly JIT compilers, to see what their inlining heuristics look like. I also read (skimmed) some papers to see what those folks had to say. I wonder if they agree. This post was a long time coming. I started working on it about five years ago but then when I quit working at Facebook I accidentally left behind all of the inliner research I did for Cinder’s inliner. So then I kind of just thought about it aimlessly for a while before redoing it this year. Anyway, here’s wonderwall. Spoiler alert: all in all, people tend to look at: And also have different interesting ways to pipe in profile information. Last, some newer papers do some wild stuff: Another thing to consider in inlining is how you gather and interpret profiles. When you compile a function, you tend to specialize it based on the input it has historically been given. For a monomorphic input, maybe you guard that the type is still the same and otherwise jump into the interpreter. For a polymorphic input, maybe you check the top K (~4) common cases and otherwise jump into the interpreter. Fine. But sometimes you can be compiling a polymorphic method that is actually monomorphic in its caller . That is, might only ever pass one kind of input to , but other callers pass all kinds of stuff. Here is a bit of a silly example to show what I mean: Just kidding, not so silly at all. It’s a super common pattern in Rails . It makes polymorphic in even though for many of its callers, it may well be monomorphic (or even a constant). In order to plumb this information through to the compiler, you have to figure out this call context relationship. There are a couple of common ways to do it. YJIT, for example, though it does not inline, splits methods based on the types of the arguments going in. This means that it clones the compiled code, generating a new version for each context. This does not give call context (“A calls B”) but gives type context (“B is called with integers, B’ is called with strings”). A compiler could do type-based splitting in the interpreter or a baseline tier. If you don’t fancy duplicating the code, you can instead duplicate the profiles. You could either do this using type context (as above) or using call context. SpiderMonkey, for example, does “trial inlining” that allows callers to pass down a bit of memory for potential inline candidate callees to record their inline caches. Instead of each function holding its own ICScript, the caller allocates a unique ICScript for that potential-inline call-site. This gives each callee function (at least?) one level of call context. Later, when inlining the callee into the caller, we don’t have other callers’ type information polluting the IR builder (or whatever reads the profiles). JavaScriptCore handles this by inlining bytecode into other bytecode. This is a gnarly transformation but gives the interpreter, even (!) access to call context. On tier-up to the compiler, all the inlining decisions have been made already. HotSpot handles this with multiple tiers. The interpreter tiers up to the client compiler, C1. C1 profiles branch and call targets in compiled code. C1 may eventually recompile based on this new information. C1 may eventually tier up to C2, which copies C1 inlining decisions. This way, we get call context in profiles via inlining. One last thing you could do is just trust your type inference and branch folding in the optimizer. You could inline and do polymorphic specialization in the callee when building the IR, then hope that your branch pruning monomorphizes the inlined callee. It’s a little wasteful because the polymorphic code is built “for nothing”, but it might work fine? Okay, onto the collected notes and half-baked commentary. Here’s a survey of a bunch of JIT compilers and how they reason about inlining heuristics. But before we get into that, thanks to Iain Ireland, CF Bolz-Tereick, and Ian Rogers for feedback on this blog post! What follows is mostly a “bits and bobbles” section a la Phil Zucker . We’ll start with Cinder , because when I wrote Cinder’s inliner I added only the simplest heuristics, mostly “don’t inline” signals. Over time, after I left, people tuned it a bit more. The inliner starts from the caller CFG, walking it to find suitable inlining candidates. Inlining candidates are only for call targets that are known—in Cinder’s case, only for monomorphic call targets—and pass some checks. The callee is only known by it’s function object, which includes its bytecode. There is no IR available for the callee until we decide to inline. Most of the “can’t handle this” checks are related to argument handling. Python has a pretty complex calling convention, so if the caller/callee have not agreed on how the arguments should be passed through, the inliner doesn’t care to try and figure it out on its own. That is the responsibility of other parts of the compiler . Things in this function could be considered “TODO”. Failures are logged so they can be analyzed. If the Cinder team determines that there is some very frequent case they should handle, they will find out from the logs. The inliner collects all candidate call instructions in one pass over the CFG. It loads the configurable “cost limit” from the options struct. Then it does one pass over the inlining candidates vector, inlining until it (maybe) hits the cost limit. It does some graph maintenance work after inlining these calls, but that’s it. This approach gets a surprising amount of utility for being so simple: it inlines constants (quite a few methods look like ), small methods, and (at least, as far as I can remember) shrinks the compiled code size. All for very little compile time overhead. There’s one other “standalone” Python JIT out there, PyPy. So we should look at that too. There are two inliners in PyPy. One is inside the RPython to C translation pipeline, which acts more like an ahead-of-time compiler 2 . Then there is the tracing JIT bit, which has its own optimizer and heuristics. We’re going to look at the latter. I talked to CF Bolz-Tereick about the inliner and their comment was that PyPy’s inlining heuristic is “yes”. There are a couple of exceptions, such as not inlining recursive functions or functions with loops. But the basic idea of tracing includes tracing through call instructions, which naturally means that you are “inlining”. PyPy also does this neat thing where they treat frame pushes like normal allocation. Frame pushes, frame reads, and frame writes get written to the trace like normal object memory traffic and can get optimized away like other field reads and writes. This means that they can “just” use DCE to eliminate frame pushes and pops, whereas Cinder has some complicated mechanism to do it (which is my fault). TODO get more details here V8 is a JS engine and it has over the years had many execution approaches. We’ll look at three of them since they all have or had their place in the history: They also each inline at different times in the pipeline, which made for a fun time trying to understand the different codebases. Inlining happens during Hydrogen graph building Don’t store function bytecode of all functions; need to re-parse callee text source to inline Heuristics https://github.com/tekknolagi/v8/blob/a969ab67f8e1e7475d9b26468225c3a772890c64/src/crankshaft/hydrogen.cc#L7807 https://docs.google.com/document/d/1VoYBhpDhJC4VlqMXCKvae-8IGuheBGxy32EOgC2LnT8/edit https://github.com/v8/v8/blob/036842f4841326130a40adfcff38f85a9b4cd30a/src/compiler/js-inlining-heuristic.h#L14 When optimizing, add call instructions to the inline candidates list: https://github.com/v8/v8/blob/1a391f98cc7a9196369f2d6cab7df35ffbe92c08/src/maglev/maglev-graph-optimizer.cc#L1271 https://github.com/v8/v8/blob/036842f4841326130a40adfcff38f85a9b4cd30a/src/maglev/maglev-inlining.h#L36 Unlike for example Cinder, Maglev looks like it does not have a lot of restrictions about what can get inlined into what, so its “can inline” signal is about budget. Actually two budgets: small budget and normal budget. Then its inlining loop is a greedy walk of the to-inline queue checking candidate sizes. It runs this loop (which drains the queue) interleaved with the optimizer (which populates the queue). Confusingly, though, the optimizer also calls another function called which checks if it legally can inline: appears unused? / dead declaration? maybe src/maglev/maglev-graph-builder.cc is just not working on github search also unused / dead declaration same JavaScriptCore is funky! Unlike these other compilers that do inlining in their neat little SSA IRs, JSC inlines at the bytecode level 4 . This is their way of making sure that they get at least one level of call context into their interpreter inline caches, which will eventually give better information to the compiler. JSC only inlines based on bytecode profile information, and only inlines bytecode?? TODO find better sources for bytecode inlining SpiderMonkey has another way of getting that call contet without doing bytecode inlining: they add call context to their inline caches. Methods can pass down an ICScript to their callees where the callee writes its inline cache information. Then, when compiling, the callee is more likely to be monomorphized. https://github.com/mozilla-firefox/firefox/blob/438a3ce10eb77fb50d968463b7741117aec5bb4a/js/src/wasm/WasmHeuristics.h#L213 SpiderMonkey ICScript https://fitzgen.com/2025/11/19/inliner.html Plan: run in interpreter; tier up to C1; profile call targets; inline in C1; profile branch counts; tier up to C2, which copies C1 inlining decisions in bytecode parser https://github.com/openjdk/jdk/blob/a05d5d2514c835f2bfeaf7a8c7df0ac241f0177f/src/hotspot/share/opto/bytecodeInfo.cpp#L116 https://github.com/openjdk/jdk/blob/497dca2549a9829530670576115bf4b8fab386b3/src/hotspot/share/opto/bytecodeInfo.cpp#L197 https://github.com/openjdk/jdk/blob/497dca2549a9829530670576115bf4b8fab386b3/src/hotspot/share/opto/parse.hpp#L42 https://github.com/openjdk/jdk/blob/497dca2549a9829530670576115bf4b8fab386b3/src/hotspot/share/opto/doCall.cpp#L185 Not too small Walk up the call stack to figure out what to compile Handling the right thing to inline: def foo(a) = a.each {|x| x } want to compile , inline each, inline block, not compile block separately (probably) https://bernsteinbear.com/assets/img/design-hotspot-client-compiler.pdf https://github.com/openjdk/jdk/blob/d854a04231a437a6af36ae65780961f40f336343/src/hotspot/share/c1/c1_GraphBuilder.cpp#L755 https://github.com/openjdk/jdk/blob/d854a04231a437a6af36ae65780961f40f336343/src/hotspot/share/c1/c1_GraphBuilder.cpp#L3854 heuristics: TruffleRuby uses weighted compile queue Graal https://ieeexplore.ieee.org/document/8661171 https://github.com/dotnet/runtime/blob/2d638dc1179164a08d9387cbe6354fe2b7e4d823/docs/design/coreclr/jit/inlining-plans.md https://github.com/dotnet/runtime/blob/0b3f3ab1ecf4de06459e5f0e2b7cb3baf70ef981/src/coreclr/jit/inline.def#L94 https://github.com/dotnet/runtime/blob/0b3f3ab1ecf4de06459e5f0e2b7cb3baf70ef981/src/coreclr/jit/inlinepolicy.cpp https://github.com/dotnet/runtime/blob/0b3f3ab1ecf4de06459e5f0e2b7cb3baf70ef981/docs/design/coreclr/jit/inline-size-estimates.md?plain=1#L5 https://github.com/dotnet/runtime/blob/0b3f3ab1ecf4de06459e5f0e2b7cb3baf70ef981/src/coreclr/jit/fginline.cpp https://github.com/dotnet/runtime/issues/10303 https://github.com/AndyAyersMS/PerformanceExplorer/blob/master/notes/notes-aug-2016.md https://github.com/dart-lang/sdk/blob/391212f3da8cc0790fc532d367549042216bd5ca/runtime/vm/compiler/backend/inliner.cc#L49 https://github.com/dart-lang/sdk/blob/391212f3da8cc0790fc532d367549042216bd5ca/runtime/vm/compiler/backend/inliner.cc#L1023 https://web.archive.org/web/20170830093403id_/https://link.springer.com/content/pdf/10.1007/978-3-540-78791-4_5.pdf An adaptive strategy for inline substitution (PDF) tracelet based https://github.com/facebook/hhvm/blob/eeba7ad1ffa372a9b8cc9d1ec7f5295d45627009/hphp/runtime/vm/jit/inlining-decider.h#L89 https://github.com/LineageOS/android_art/blob/8ce603e0c68899bdfbc9cd4c50dcc65bbf777982/compiler/optimizing/inliner.h https://github.com/JikesRVM/JikesRVM/blob/5072f19761115d987b6ee162f49a03522d36c697/rvm/src/org/jikesrvm/compilers/opt/inlining/DefaultInlineOracle.java#L55 Partial inlining Understanding and Exploiting Optimal Function Inlining (PDF) machine learning Automatic construction of inlining heuristics using machine learning Machine-Learning-Based Optimization Heuristics in Dynamic Compilers (PDF) Guiding Inlining Decisions Using Post-Inlining Transformations (PDF) U Can’t Inline This! (PDF) Towards better inlining decisions using inlining trials RhizomeRuby inlining An Optimization-Driven Incremental Inline Substitution Algorithm for Just-in-Time Compilers (PDF) Automatic Tuning of Inlining Heuristics (PDF) Inlining-Benefit Prediction with Interprocedural Partial Escape Analysis (PDF) Inlining of Virtual Methods (PDF) A Study of Type Analysis for Speculative Method Inlining in a JIT Environment (PDF) A Comparative Study of Static and Profile-Based Heuristics for Inlining (PDF) clusters from Custom benefit-driven inliner in Falcon JIT (PDF) https://github.com/oracle/graal/blob/5dde777cba22a99ebe3f19745d03ddfbc35c563c/compiler/src/jdk.graal.compiler/src/jdk/graal/compiler/phases/common/inlining/policy/GreedyInliningPolicy.java https://github.com/oracle/graal/blob/5dde777cba22a99ebe3f19745d03ddfbc35c563c/compiler/src/jdk.graal.compiler/src/jdk/graal/compiler/phases/common/inlining/InliningPhase.java https://github.com/oracle/graal/blob/5dde777cba22a99ebe3f19745d03ddfbc35c563c/compiler/src/jdk.graal.compiler/src/jdk/graal/compiler/phases/common/inlining/info/elem/InlineableGraph.java#L148 There are some newer papers, especially in Java land, that try to do a lot of analysis ahead-of-time and bundle the resulting information in .class files. Then the JIT can read it and see more than local context. Or, if you are an AOT compiler, you can probably do a lot more whole system reasoning—both for time budget reasons and also because you can see more functions at once.  ↩ Check it out if you like. I stumbled across it by accident.  ↩ See also “Turbolev”, which seems to merge Maglev (CFG) with Turbofan (Sea of Nodes)… somehow.  ↩ Potentially a misunderstanding based on a private conversation. I’m working on tracking down the implementation…  ↩ Profiles of call target Cumulative caller size (increasing as callees get inlined) Callee size Inline depth Number of inlined calls at a certain depth If recursion is present Callee/caller call count ratio (if callee only called less than K% of calls to caller, don’t inline callee) Callee stack usage Polymorphism in callee What mode the compiler is in (baseline vs more aggressive) If the callee looks like it always raises/throws Train neural networks to make inlining decisions Let inlining drive the entire optimization pipeline, treating it as a search heuristic over a BFS walk of the call graph Use AOT-gathered information to aid in JIT heuristics Hydrogen was the first real SSA IR and it looks very familiar to me, having worked on Cinder and now ZJIT. It is now defunct. Turbofan was the replacement, going full Sea of Nodes. In the grand scheme of things it is a pretty fast compiler, but it does not hold back from doing some expensive rewrites. This was recently rewritten from Sea of Nodes to a mode traditional CFG and nicknamed Turboshaft. Maglev is meant to coexist alongside Turbofan, preferring to speculate a little more eagerly and do fewer incremental rewrites in the name of compile time. 3 https://github.com/tekknolagi/v8/blob/a969ab67f8e1e7475d9b26468225c3a772890c64/src/crankshaft/hydrogen.cc#L9236 something about native context check callee AST size against configurable limit check inlining depth against configurable limit don’t inline recursive functions check current cumulative method size (as tracked by AST node count) against configurable limit Find candidates https://github.com/v8/v8/blob/036842f4841326130a40adfcff38f85a9b4cd30a/src/compiler/js-inlining-heuristic.cc#L134 Can inline https://github.com/v8/v8/blob/036842f4841326130a40adfcff38f85a9b4cd30a/src/compiler/js-inlining-heuristic.cc#L75 Force inline small functions https://github.com/v8/v8/blob/036842f4841326130a40adfcff38f85a9b4cd30a/src/compiler/js-inlining-heuristic.cc#L309 Loop over sorted (by comparator) list https://github.com/v8/v8/blob/036842f4841326130a40adfcff38f85a9b4cd30a/src/compiler/js-inlining-heuristic.cc#L847 skip recursion https://github.com/v8/v8/blob/1a391f98cc7a9196369f2d6cab7df35ffbe92c08/src/objects/shared-function-info-inl.h#L421 not called enough (min call frequency) bytecode too big Bytecode inlining https://github.com/WebKit/WebKit/blob/709c3895afd71e0836f8c8be7393e44d41fab7e1/Source/JavaScriptCore/bytecode/CodeBlock.cpp#L2453 DFG https://github.com/WebKit/WebKit/blob/709c3895afd71e0836f8c8be7393e44d41fab7e1/Source/JavaScriptCore/dfg/DFGCapabilities.cpp#L76 https://github.com/WebKit/WebKit/blob/917854a9c245b87b333e23ed4b195505d574a333/Source/JavaScriptCore/dfg/DFGByteCodeParser.cpp#L1703 https://github.com/WebKit/WebKit/blob/917854a9c245b87b333e23ed4b195505d574a333/Source/JavaScriptCore/bytecode/CallLinkStatus.cpp#L294 https://github.com/WebKit/WebKit/blob/d919344236c47b610930636d3310f00380624d43/Source/JavaScriptCore/bytecode/InlineCallFrame.h skip callees with exception handlers (unless explicitly allowed with a CLI flag) skip synchronized callees (unless explicitly allowed with a CLI flag) skip classes with unlinked callees skip uninitialized classes max inline level (default 9) max recursive inline level (default 1) callee bytecode size (max for top level is 35 bytecodes, but falls off by 10% per inline level) callee stack usage (max of 10 slots) max total method size (default 8000 bytecodes) There are some newer papers, especially in Java land, that try to do a lot of analysis ahead-of-time and bundle the resulting information in .class files. Then the JIT can read it and see more than local context. Or, if you are an AOT compiler, you can probably do a lot more whole system reasoning—both for time budget reasons and also because you can see more functions at once.  ↩ Check it out if you like. I stumbled across it by accident.  ↩ See also “Turbolev”, which seems to merge Maglev (CFG) with Turbofan (Sea of Nodes)… somehow.  ↩ Potentially a misunderstanding based on a private conversation. I’m working on tracking down the implementation…  ↩

Dart Performance C# C++ Programming

▲ 0

0 views

Oya Studio 1 months ago

Rive is the server-driven UI engine I wanted

Rive is sold as a tool for cute animations. The more I use it, the more I see something else entirely: a genuinely cross-platform engine for server-driven UI. Here is how to load a component from a remote .riv file and bind it to dynamic, per-locale data in Flutter.

Mobile Flutter Dart

▲ 0

0 views

Ed Zitron's Where's Your Ed At 1 months ago

AI Is Too Expensive

If you liked this piece, you should subscribe to my premium newsletter. It’s $70 a year, or $7 a month, and in return you get a weekly newsletter that’s usually anywhere from 5,000 to 18,000 words, including vast, detailed analyses of NVIDIA , Anthropic and OpenAI’s finances , and the AI bubble writ large . My Hater's Guides To Private Credit and Private Equity are essential to understanding our current financial system, and my guide to how OpenAI Kills Oracle pairs nicely with my Hater's Guide To Oracle . This week, I’ll publish the second part to my ongoing series (“ What If…We’re In An AI Bubble? ”) about the factors and events that will cause the AI bubble to finally pop.  Subscribing to premium is both great value and makes it possible to write these large, deeply-researched free pieces every week.  AI is, as it stands, not economically viable for anybody involved other than the construction firms, NVIDIA, and the surrounding hardware companies benefitting from the irrational exuberance of a data center buildout that doesn’t appear to be happening at the speed we believed .  Every AI startup loses millions or billions of dollars a year, and nobody appears to have worked out a way to stop hemorrhaging cash. Hyperscalers have invested over $800 billion in the last three years, with plans to add another $700 billion or so in 2026 and another $1 trillion in 2027 , meaning that they need to make at least three trillion dollars in AI specific revenue just to break even , and $6 trillion or more for AI to be anything other than a wash. I went into detail about this (albeit at a lower, pre-2026/2027 capex number) in a premium piece last year .  To give you some context, Microsoft made $281 billion , Meta $200 billion , Amazon $716 billion , and Google $402.8 billion in revenue in their most-recent fiscal years for every single product combined, for a total of $1.599 trillion. None of them will talk about their actual AI revenues. Yes, yes, I know Microsoft said that it had $37 billion in AI revenue run rate ($3.08 billion a month or so) and Amazon had $15 billion, or around $1.25 billion a month , but both of these are snapshots of single months that are meant to make it sound like they’re going to make that much in a year but in the end, you don’t actually know anything about how much money they’ve made from AI. We do, however, now know that Microsoft has spent an approximate $100 billion on its OpenAI partnership after testimony from an executive during the otherwise-dull Musk-OpenAI trial, per Bloomberg : This is a fascinating insight for a few reasons: At the end of 2025, OpenAI claimed that it had 1.9GW of capacity (likely referring to total power draw rather than the actual critical IT of the infrastructure at its disposal), which, per analyst estimates, ( $42 to $44 million per megawatt ) works out to around $79.8 billion. This claim was made around six months before the release of Microsoft’s most recent quarterly results.  In other words, Microsoft has spent 4 years sinking (either through spending or allocating the capex in advance) nearly $300 billion into…building OpenAI? Okay, fine. Microsoft also has 20 million Microsoft 365 Copilot subscribers for an absolute maximum revenue of $7.2 billion…if every single one were paying $30 a month, which they are most assuredly not as Microsoft has been offering discounts on it for years . Based on my reporting from last year , Microsoft made around $7.5 billion from OpenAI’s inference spend and $761 million from its revenue share in Fiscal Year 2025, a year when it invested (either spent or allocated) around $88.2 billion in capital expenditures. I didn’t report it at the time, but I also had the numbers for all of Microsoft’s revenues for the first three quarters of Fiscal Year 2025 — a total of $8.9 billion of total AI revenue, with around $4.35 billion in revenues when you removed OpenAI’s inference. If we assume that Microsoft’s other AI services grew 10% quarter-over-quarter, I estimate that Microsoft likely made around $17.9 billion in AI revenue in FY2025, or a little under a fifth of its capex.  And let’s be clear: none of these numbers include the actual operating expenses. Data centers, after all, need electricity to run, and AI data centers in particular need a lot of electricity. And some — though, admittedly, not many — people to handle the things like maintenance, repairs, and operations. And then there are things like taxes, insurance, and the other day-to-day costs that, when you add them all together, make a big, scary number.  You can argue that “actually GPUs are profitable to run” ( I disagree! ), but for any of this to make sense, four things have to happen: All four must be true. If AI revenues don’t explode, capex can stop, margins can be positive, and your best-case scenario is…you maybe broke even. If capex never stops being invested, you need revenues to explode dramatically — to the tune of effectively doubling Microsoft, Meta and Google’s entire businesses, and tripling Amazon Web Services’ annual revenue ( $128 billion ) — and for said revenues to be margin-positive, because if they’re not, eventually other healthy businesses will slow, leaving AI to tear a hole in overall margins. In all cases, AI revenue must stay consistent because, well, you need to get paid . I also cannot find an economic scenario where this pays itself off.  Let’s assume that Anthropic is actually at $45 billion in annualized revenue ( I believe it’s doing some very worrisome maths to get there ), or around $3.75 billion a month. On an annualized basis, this would not be enough — assuming it had zero operating expenses (rather than losing billions) — to recover a single year of capital expenditures from Microsoft, Google, Meta, or Amazon from 2024 or 2023. Even if OpenAI’s entire cloud spend ( $50 billion ) for 2026 went to Microsoft and it doubled its Microsoft 365 Copilot revenue (at full cost) to $14.4 billion, it estimates it will invest $190 billion in capital expenditures this year. Amazon’s $15 billion AI run rate, even if it doubled, wouldn’t put much of a dent in its $200 billion in investment plans . While we don’t know Google’s AI revenues, it plans to invest $185 billion in capex this year . These AI revenues have to be completely fucking insane and they need to be that way extremely fucking soon , because otherwise the best they’ll be able to say is “our first few years of capex weren’t particularly useful but the stuff we built after it was,” which still works out to a few hundred billion dollars of waste. Things get even worse when you realize that at least 70% of Microsoft, Google, and Amazon’s compute is dedicated to Anthropic and OpenAI , two companies that burn so many billions of dollars that Microsoft, Google and Amazon have already fed them a combined $54 billion in the last three years, with $28 billion of that coming in the last month and Anthropic due another $50 billion from Google and Amazon if certain performance obligations are met. And there’s no real sign, outside of Anthropic and OpenAI’s compute spend (which is reliant on hyperscaler and venture capital money), of any real explosion in AI revenue. Per The Information (in a chart I love to share!), more than 50% of hyperscalers’ revenue backlogs comes from these companies: If massive, incredible demand for AI existed, wouldn’t these remaining performance obligations be near the trillion mark? Wouldn’t there be other Anthropic or OpenAI sized chunks of revenue? There’s allegedly incredible, unstoppable, insatiable demand for compute. Why isn’t it lining up? Let’s take a look at those RPOs! That was a lot of numbers, so let me make it simpler: outside of OpenAI and Anthropic, these three companies do not appear to be significantly increasing their revenues, and the only way to get that revenue is to feed money to one or both of these companies.   Put aside all the theoreticals and hypotheticals and metaphors and imaginary future scenarios and tell me: what, in the next year, are Microsoft, Google and Amazon going to do about this problem? How do they solve it? If we assume the absolute best-case scenario, these companies are making a combined $70 billion in annual revenue on investments that now — including the money invested in the companies themselves — total over $900 billion. Doubling that won’t be enough. Tripling it won’t be enough. In fact, to pay this off, these companies will need to be making over $100 billion each in AI revenue in the next year , because otherwise there is no covering these losses. And it all comes back to a very simple point: AI is too expensive. If the margins were good, they’d be sharing the margins. If the revenues were good, they’d be sharing the revenues (and no, run rates aren’t revenues). If the business was strong, it would be a separate category in their earnings.  But LLMs are too expensive! They cost too much to run, and said costs appear to increase linearly with revenues. The more a user uses a product, the more it costs the company to run it, and the more capacity they can take up. The only way to capture any growth is to buy and install GPUs , which in turn requires you to build somewhere to put them, which takes time and money.  I’m really struggling to see the argument in favor of continued capex investment. You’re more than $800 billion in the hole with, I estimate, less than half of that resulting in operational GPUs and capacity. Said capacity is mostly taken up by OpenAI and Anthropic, two companies that burn billions of dollars and do not appear to have an answer for how they might stop.  The more you build, the more your infrastructure becomes dependent on the continued existence of two perennially-unprofitable ultra-oafs, as your existent AI product lines are, at best, add-ons to products like Google Workspace or Microsoft 365, or further expansion of cloud compute capacity with lower margins and higher up-front costs than anything you’ve ever built.  Every quarter is an opportunity to put yourself another $30 billion or so in the hole, all in the hopes that, I assume, OpenAI or Anthropic will pay you $100 billion or $200 billion over the course of a few years, because nobody else in the entire universe is spending that much on compute. You are not recovering these investments without either a massive new product line that doesn’t exist today or three or four Anthropic or OpenAI-sized compute contracts. Put another way, Amazon needs another AWS ($128 billion a year), Microsoft another Azure ( $75 billion a year, including OpenAI’s 2025 compute spend ) and Google a business line at least half the size of search (around $200 billion a year). These businesses have grown to this size by providing extranormally large amounts of value from the very moment they were created and impenetrable monopolies — and while there are quite literally other cloud providers that can physically provide the infrastructure to OpenAI and Anthropic ( Oracle is trying to compete and may die as a result ), the actual “monopoly” here is “being able to deploy hundreds of billions of dollars.” Anthropic proved this when it took 300MW of compute from Elon Musk .  In Oracle’s case, as I’ve explained at length , it has to successfully build 7.1GW of capacity, have that capacity actually be margin-positive (doubtful!), and then actually get paid for it by the time it’s built in, oh, I dunno, 2032?  Sadly, I have bad news about Oracle, Microsoft, Amazon, and Google’s largest customers.  Here’s a fun game: ask an AI booster how OpenAI or Anthropic becomes profitable! Here’s what they’ll say: I must be abundantly clear that nobody has any proof that anyone is profitable on inference, but we have plenty of proof they’re not. They’ll likely cite known liar Sam Altman saying OpenAI is profitable on inference from a party from August 2025 , or Dario Amodei saying ( in a sentence around “stylized facts” that are “not exact” and are specifically “a toy model” and specifically not about Anthropic ) “the inference has some gross margin that’s more than 50%.”  Here’s a really simple way to dispute this: Coatue said that Anthropic’s revenues were 85% API calls in 2025 . If it’s profitable on inference, how is it still losing money? You’re gonna say “training,” but that doesn’t actually answer the question: if Anthropic’s process of providing tokens to its models is profitable, how is it losing so much money? Why offer a subscription platform at all?  As I’ll get to, Anthropic has companies paying massive amounts for tokens — hundreds of millions a year in some cases — that’s all inference . Why are you bothering with these stinky, nasty monthly subscriptions? The “inference is profitable” argument is a bedtime story told to people that can’t reconcile the logic of a company that allows people to burn between $8 and $13.50 of every dollar of their subscription revenue.   Otherwise, you have to reconcile with the fact that both Anthropic and OpenAI are both incinerating money and have no real path to any kind of sustainability other than, well, not doing that. One very, very specific counter-argument people make is that open source models are cheap, and can somehow be compared to OpenAI and Anthropic’s, despite the fact that we have no idea what the actual parameters of Sonnet, GPT, Opus, or any other of their models actually are.  What we do know is that both of these companies lose billions of dollars. What we do know is that OpenAI, per The Information , plans to burn $852 billion through the end of 2030, and that as of March 6, 2026 (per CFO Krishna Rao’s sworn affidavit), Anthropic made “exceeding” (sigh) $5 billion in revenue and spent $10 billion on inference and training.  Anthropic has done a great deal of work to obfuscate how much it actually makes or spends, but I think it’s likely it burns even more than OpenAI, given the fact that it’s had to raise $75 billion in the last 6 months ( assuming its new $30 billion round closes ), and that’s not including an additional $30 billion from Google and Amazon if certain unknown milestones are hit.  Then there’s the issue of those RPOs. Anthropic is now on the hook for $200 billion to Google, $100 billion to Amazon and $30 billion to Microsoft, I assume over the course of the next three or four years.  So let’s lay this out. Anthropic — based on its own affidavit from March — appears to have spent $3 to make $1 of revenue on a compute basis, and that’s before you include any and all other costs like staff or electricity or the vocal coach that Dario Amodei uses to add that bass to his voice.  Additionally, it needs $330 billion to pay its cloud obligations to Amazon, Google, and Microsoft over the next four years. I’d estimate it needs $5 billion a year for its compute deal xAI (so $20 billion over the total period) and an estimated $30 billion to cover its deal with CoreWeave . That brings us to a total of $380 billion. It’s hard to estimate the actual costs associated with running Anthropic because so much of the reporting no longer makes sense as a result of that affidavit. Nevertheless, I think it’s fair to assume it will need at least $20 billion of operating expenses across that four year period. We don’t even need to play in the realm of “what might Anthropic or OpenAI’s revenues be?” to understand the problem here. Both companies aggressively burn money, and neither of them have any answer as to how they might stop. Numerous reports about how Anthropic will turn “cash flow positive” in either 2027 or 2028 are fantastical, illogical, entirely driven by ridiculous projections, and should never have been reported as anything other than an attempt by companies to mislead their investors. In both cases, reporters should’ve had more asterisks on those numbers than Q*Bert reading Frank’s lines from Blue Velvet . And we have plenty of evidence that they’re losing more money over time. In January 2026, The Information reported that Anthropic’s gross margins were 40% in 2025 — 10% lower than its “optimistic” projections, specifically attributed to “...the costs of running Anthropic models from paying customers, in a process known as inference, on servers from Google and Amazon,” adding that those costs were “23% higher than the company anticipated.” In February, The Information ran another story saying that OpenAI’s gross margins fell from 40% in 2024 to 33% in 2025, a full 13% lower than its projected margins of 46%, all because (and I quote) “...the company having to buy more expensive compute at the last minute in response to higher than expected demand for its chatbots and models.” You know, exactly what Anthropic has had to do. This is what I’ve referred to as the knife-catching problem for compute demand — you either don’t order enough compute and have to rush to buy some last-minute as demand intensifies, or you order too much, and, well, to quote Dario Amodei: And right now, as I’ve covered , there’s not enough compute being built to keep up with Anthropic or OpenAI’s voracious demands, meaning that they will both be bartering to buy whatever’s available at whatever price it’s available at. This naturally will savage their already-negative margins… …and then what? No, really, and then what? One of you fucking AI boosters, answer me, how does this actual reverse course? Because even if Anthropic were making $100 billion in annual revenue, it would probably be losing $300 billion or more to get there. The fact it had to raise $30 billion in February , $15 billion in April, and now $30 billion more in May all while allegedly pulling in more than $3 billion a month in revenue suggests that its COGS are fucking horrendous, and its growth is coming at a terrible financial cost. Let’s say that Anthropic keeps growing and ( as The Information suggests ) hits $100 billion in annualized revenue (around $8.3 billion a month). How, exactly, does it afford to make that much money? Because right now it’s (allegedly) about to hit $45 billion in annualized revenue, and needs so much money that it’s absorbing (along with OpenAI) the majority of venture capital raised this year, and very clearly does not have any path to bring its costs down. The answer is simple: it can’t! There is no mechanism to do so. More compute does not make OpenAI or Anthropic’s services cheaper to offer. There is no magical silicon coming that will make any of this more affordable, and no, Anthropic is not “profitable on inference,” because if it were, that massive revenue growth would have leveled out its margins rather than require it to raise a little less than the combined value of every Major League Baseball team , or more if you add the other $50 billion that Amazon and Google have promised based on secretly-held performance obligations. The same goes for OpenAI, which “raised” $122 billion (around $45 to $50 billion in real cash, with the rest either paid in installments or on it IPOing or reaching (sigh) AGI) in February and is now already considering raising more . Somebody might counter-argue that this is companies raising as a means of boosting their valuations, I think that’s a very convenient way of looking at two extremely problematic companies.  I should also ask why neither of them appear to be seriously considering going public. While both were rumoured earlier in the year to be planning to do so in 2026, both appear poised to raise more private capital. I think the answer is simple: their CFOs know that doing so would reveal their actual margins, which are hot dogshit with sprinkles on top.  Nobody has a sensible or logical response here. Which leads us right to our next point! One important detail to keep in mind here is that as of a month or two ago, Anthropic moved all enterprise customers to token-based-billing, which will begin, I believe, a true stress-test of the true “value” of AI as costs skyrocket. Just last week I ran the first of a two (or three, potentially) part premium series called “What If We’re In An AI Bubble?” and touched on the gruesome subject of whether organizations could afford to pay for AI long-term : Earlier in the week, carnival barker and Salesforce CEO Marc Benioff said his company would spend $300 million on Anthropic tokens in 2026 , and as I discussed in my premium from Friday , unrestrained AI spending is inflating the revenues of Anthropic and OpenAI in a way that isn’t sustainable for anybody involved: The problem is simple: nobody actually knows how much AI is going to cost them in any given quarter. This means that the current token spend you’re seeing is entirely experimental, which is why organizations keep burning through their tokens so fast.  This massive growth in spend is what underpins the “massive” (I have serious questions about its accounting) growth in Anthropic’s revenue. Executives have, across the board, given their engineers free reign to burn as many tokens as they’d like, and while I severely doubt that Anthropic actually hit $50 billion in annualized revenue outside of not-quite-fraudulent non-GAAP measurements, I believe its revenue growth has come from an artificial boost from a tech industry searching for a reason to pay somebody money. To be very clear about what I mean, I think there is currently an AI token binge across both Anthropic and OpenAI. Enterprises do not know the actual value of AI, and do not know how much they should actually be budgeting, which is why Uber and others are running through their token budgets but not, it seems, spending less. We’re currently in an abundance phase — one where nobody is truly thinking about the costs outside of their fear of missing out — but there’s this nasty undercurrent of “wait, how much does that cost?” followed by “oh, fuck, well…you know I love AI but…” Put another way, the current spend on AI tokens is not something that’s indicative of lasting, reliable revenue. In some cases, the pressure to use AI for everything is turning companies’ software stacks into slop. Things are worse elsewhere. Something is wrong at Zillow. Something about LLMs has done something to its technical leadership, something that makes them talk strange and send weird slide decks with confusing, slop-ridden sentences.  The real estate tech firm spent over $1 million on AI services in the first quarter of 2026, and in April it spent $749,000 in tokens across Cursor and Anthropic’s services, as well as through AWS Bedrock. As of the end of the month, it was nearly 75% of the way through its annual Cursor token budget of $1.1 million.  As of the middle of May, its total AI spend had already crested over $300,000, and its Cursor budget sat dangerously close to the edge at 85%. This is particularly-concerning when you consider that Zillow’s net income for Q1 2026 was $46 million , and ranged from $2 million to $10 million each quarter of 2025.  Zillow is currently on course to spend at least $7 million on AI in 2026, and at its current pace might hit as much as $10 million, which would amount to a little less than 50% of its 2025 net income ( $23 million ).  You’re probably wondering how Zillow manages to spend so much on AI, and the answer — as I’ll get into in next week’s free newsletter — is that its technical executives appear to have AI psychosis, saying that the short-term goal is for “software engineers to never open a code editor again.” The reality is chaos. In a slide deck that I’ll discuss later, Zillow revealed that while engineering resources have largely stayed the same, outputs requiring human review have increased by nearly 50%. Meanwhile, code deployments and pull requests increased by 39%, and software reviewer load increased by 29,000 hours each month , creating a massive burden on the 1,500 or so engineers working at the company.  In simpler terms, that’s about 19 hours of extra work per engineer that’s literally just looking at extra code written by LLMs.  On Blind, the anonymous social network for tech workers, Zillow workers complain about Zillow’s code “slowly becoming AI slop,” with “much more code getting approved without guardrails or input due to people not being able to keep up the other’s velocity or just not caring anymore.”  One worker claimed that “the slop is job security,” adding that they “don’t want the output to be good or documentation to be clean [as] management will replace [them] with offshore/nearshore/AI agents at the slightest whiff of evidence that the slop cannon is self sustaining.” Another said that they felt “lost in the agentic world,” and that they “didn’t have full grasp of where we are going or what [their] role is,” with a “lot of overlap in what people are doing.” Another said that “people are burning tokens just to hit internal AI adoption targets,” adding that “this is what happens when leadership ties metrics to usage instead of outcomes,” saying that it “literally subsidized busywork.” This is all part of what an internal slide deck viewed by this publication called “AI-Native Engineering,” promising a “path to an agentic Zillow” and “faster outcomes for customers,” though customers are never mentioned in any other slide.  The deck — pumped full of AI-generated text — talks about “generic AI being a commodity,” saying that “Zillow-aware AI is a competitive advantage,” and at no point explains what that means. It encourages engineers to go from “AI-Assisted” to “AI-Native,” with “systems enabling org-wide leverage,” with engineers moving from being “soloists” — individual developers with AI tools — to “conductors” that orchestrate AI agents, to “composers” that “define systems AI can safely play,” adding that “2026 is the transition from conductor to composer.” Yet the strangest part is named “2027: A Tuesday,” discussing a theoretical day in the office for whoever is left at the company. This theoretical example is, apparently, a process that would take weeks, but now takes under two hours.”  Zillow intends, based on this deck, to sacrifice everything to AI — code review, vulnerability fixes, policy checks, deployments, testing, and basically having agents take over everything , no matter how small, like having an agent do dependency updates and security hotfixes that could be handled with a simple shell script. To quote Zillow: In practice, sources at Zillow tell me that there has been no actual movement toward this vision. Software engineers still open IDEs and review code manually, with one describing Zillow’s “vision” as “nonsense,” adding that “you can’t just throw buzzwords on a slide deck and change how all the engineers do their jobs.”  As for why token burn is so high, sources tell me that engineers are actively encouraged to use AI for everything , as much as possible, writing PRD (product requirement documents) in AI, then using the AI to make stuff based on the PRD, then doing a deck with AI, then writing emails with AI, using AI to brainstorm, or create weird, esoteric automations, with some managers pushing workers to have one personal AI “goal” to aspire to. Zillow’s agentic “vision” is apparently a remit from the C-suite. It’s hard to tell if this is AI psychosis or just classic Business Idiot bullshit.  Perhaps it’s a little of both. Every organization I’ve talked to has exceeded or is nearing the edge of their annual token budget barely five months into the year, which means that everybody has suddenly given themselves an extra few million dollars’ worth of operating expenses for reasons that escape effectively everybody I’ve talked to.  Every engineer tells me the same thing: “I’m being made to do this, I don’t want to do this, my managers do not seem to understand, my bosses seem to understand even less than my managers, and if I don’t use AI somebody is going to fire me.”  Put another way, CEOs and CTOs are screeching at their underlings to “use AI as much as possible” to “find its incredible benefits” without anybody really knowing what those are and how much it’ll cost to get there. This might be because Anthropic obfuscates the data that might tell customers the real costs.  Per Laura Bratton at The Information , Bratton’s article has numerous quotes from executives saying that Anthropic lacks transparency and granularity into the ways that tokens are being burned across an organization, in a way that I think sounds very, very suspicious, particularly when you add the following:  While I’m not accusing Anthropic of anything untoward, massive, multi-million dollar contracts that involve individuals burning thousands or tens of thousands of dollars’ worth of tokens with no service level agreement, transparency or true granularity into the burn is a perfect setup for a company — not saying it’s Anthropic! — to do something dastardly with those numbers.  While an individual might be able to monitor their own personal usage, in an organization of hundreds or thousands of engineers, who’s to know if, say, the particular token burn is consistent across every member of the company, or that those costs are actually matching up with what the user is doing? This is a company ostensibly worth $900 billion dollars acting with disregard for the basic measurement of “how much did this cost, and how did it cost so much?” And in the end, how do you even measure it at scale? Say you’ve got 1,500 engineers, and they’re spending a combined $1 million tokens a month. How the fuck do you actually measure the return on investment for that spend?   How many tokens does it take to do one thing? Is it consistent across every model? Is it consistent across every employee? Are you even measuring how many tokens a task costs? Because if you’re not, that token budget is basically throwing a dart blindfolded.  Okay, now you’ve measured a task, did you make sure to measure it multiple times? Because LLMs can randomly do things differently even with the same prompt and same Claude.MD file and same strictures and same data sources. You’re gonna need at least 10 samples of each task, and you’re gonna need to make sure somebody who actually knows what they’re doing can measure them, because if you get a dimwit, they’re going to say it can do something it can’t. Unless, of course, you can’t actually measure how many tokens a particular task can take with much accuracy, in which case every single AI token budget is bullshit. And each model does things differently depending on many different variables, some of them a result of the user, some of them a result of the AI labs themselves. Alright, well, maybe you just need KPIs — measurements you can aspire toward , and by pursuing them you can start working out how much it costs to do stuff.  Wait, which metric works there exactly?  In fact, it’s pretty hard to measure anything like “efficiency” or “productivity” in any business, because every metric connected to them can be gamed, leaving managers and executives with the problematic situation where they have to start learning how things work so they can see if they’re good. Before AI, this wasn’t as much of a problem, in the sense that inefficiencies and wasted hours weren’t directly connected to a chatbot that is specifically designed to burn money. Managers and executives could come up with whatever deranged, self-gratifying office bullshit they pleased, wasting hours of people’s time in the process, but doing so didn’t immediately connect to a massive, ever-increasing cost. AI is a perfect storm of failed concepts and organizations, and the apex of the Era of the Business Idiot , an epoch where we’re ruled by people so thoroughly disconnected from the actual workforce that it was inevitable that a technology would be created specifically to grift them. LLMs are dangerous for many, many reasons, but the under-discussed one is how well they play to a certain kind of executive imbecile. Generative AI is — to quote Mo Bitar — really good at doing an impression of work, much like most managers and c-suite executives, and even if it’s completely incapable of doing something, it’ll absolutely say it can and tell you you’re amazing for suggesting it. And that’s why Business Idiots love it.  Where regular human beings would say annoying things like “that’s not possible within that timeline” or “we don’t have the resources to do it,” AI will say “of course, right away!” and burn as many tokens as possible.  When it makes mistakes, it’ll apologize — as it should because it failed you — but then promise to do better next time, all while costing so much less, at least in theory , than a regular, stinky human being.  It’ll create a PRD of a theoretical software project with the confident and vigor that you need to take it immediately to a software engineer and say “build this immediately,” and when the software engineer tells you a bunch of bullshit about it not being possible , it’ll spit out several convincing-sounding responses. Fuck, why even bother talking to that engineer at all? Claude Code can mock up a prototype that you can then shove in their fucking face before you fire them for not using AI to do it themselves. Any executive-level fuckwit you’ve met in your life now has a seemingly-powerful tool that can burp up mimicry of open source software and, if you constantly prompt it, eventually get something half-functional onto some sort of web server. When you face bugs, it’ll try and fix them, sometimes also “fixing” (adding or deleting code) from elsewhere to be helpful, like when Cursor using Anthropic’s Claude Opus 4.6 model deleted an entire production database and all its backups . It will never, ever say no, even if it’s incapable, even if it has no thoughts, even if what you are asking is equal parts impossible and unreasonable in both its timescale and scope. A Business Idiot, given his druthers, can sit there and fuck around and make an LLM spit out something that makes him feel like he’s coding, which in turn makes him feel that you, a lazy and stupid engineer , could do even more with the power of AI. It doesn’t matter that it costs an absolute shit-ton of money, or that there’s no way to measure its efficacy. The Lion does not concern himself with things like “efficacy” or “productivity,” and the Lion is increasingly tired of your whining! The Lion doesn’t even understand what it is you do every day other than not doing what The Lion is asking for! You laugh, but this is genuinely how the majority of managers and executives think and act, and now they have a special chatbot that can fart out functional-enough prototypes to convince a Business Idiot they can do anything, because executives and managers do not regularly do much work and thus have no idea what it looks like other than when they look over your shoulder, which is why they wanted you back in the office! Organizations aren’t burning millions or hundreds of millions of dollars a year on AI because it’s good , they’re doing it because they are run by people who do not know what the fuck they’re doing.   In a sane world, randomly adding a massive, ever-expanding operating expense to your business with the express intent of — to quote IT firm Workato’s CIO , “eating the costs while employees experiment” — would have the board blow up your house. In our world, one dominated by disconnected, self-involved and massively-overpaid dullards, many businesses pushing their workers to use AI are doing so because the other guy is doing it, with about as much strategy and forethought as one would expect from somebody who spends 90% of their life reading emails, going to meetings, or going to lunch. The majority of those I see trumpeting the so-called benefits of AI do not appear to do anything of note. I have yet to see one so-called multi-agent orchestrator engineer psychopath ship something remarkable or impressive or even functional. I have yet to see any AI-obsessed boss write or create or author or do anything I can remember. I don’t see any of these fuckwits running a company on their own outside of those who have learned to sell stuff to other AI psychosis victims or executive midwits of varying size.  And why oh why is it always the language of inevitability and possessiveness? Nobody who’s this insistent, aggressive and violative with their language of “it’s here and if you don’t adopt it you’re stupid and dead” has ever been right about anything. Nobody this desperate, insistent and forceful has ever had good intentions, good vibes or brought good omens — they are always bearers of some kind of con.  Most technology is sold on elevating and ascending human beings. AI cheapens every interaction by creating a work-shaped product from a person that doesn’t respect you enough to give you work that’s barely fit for a human because it wasn’t made for one.  You must accept becoming a dogshit dealer that loves accepting and receiving low quality goods. You must celebrate intentionless and decaying slop, and defend it and the machine that made it with your entire being. You must sully yourself — treat its unexceptional, sloppy and unreliable outputs as signs of sentience, or at least the proof that digital sentience is possible. You must defend horrible, abrasive, ugly, loud monoliths of steel full of $50,000 graphics cards. You must say they are necessary, and you must aggressively antagonize those who do not.  Every time you defend generative AI you defend a machine of capital that has burned $1 trillion and created one of the most-wasteful products in history. If people disagree with you, you must attempt to harm them somehow — ostracize them, mock them, attack them, denigrate them. You will justify this as moral, because you have been manipulated by a technology built and sold by two of the greatest grifters of all time — Dario Amodei and Sam Altman.  Anything less is opposition to an industry with all the trappings of authoritarianism down to the media toadies, the propaganda and the seizure of land in the name of a nebulous “greater good.” But man, these men got people good.  Sam Altman helped propagate a technology perfect for conning people with potential, a larger extrapolation of Altman’s own life of taking dogshit — Loopt, for example! — and parlaying it into larger opportunities. It can make a really half-hearted demo of a lot of things, and that’s good enough to sell to Business Idiot.  Dario Amodei took this grift and perfected it. Anthropic is a company purpose-built to con people into giving it by money by making people feel smart. LLMs can do work-shaped stuff, sometimes, as long as you debase yourself to accept mediocre and often-broken stuff that you have to keep a vigilant eye on, and either use a subsided product that loses Anthropic money or pay a shit ton of money as an enterprise to Anthropic and they still lose money.  These companies were only capable of growing in an economy dominated by the gullible and work-shy. Only a capitalist culture dominated by people who don’t actually do or know stuff have let this get so far. Nobody wants this, nobody wanted it since the beginning, it was forced upon everyone, and to pretend otherwise is laughable and offensive. The amount of people who use this shit a bit and become convinced that we’re mere years from it costing over a trillion dollars to somehow making trillions of dollars and being an entirely different and good product should be aware that they are being manipulated. The more you feel compelled to defend AI the more scrutiny you must show it.  I am not your enemy! If you think that I am, you are on the side of a corporation or a product. You can try it, like it, and I don’t really care, but the second I see you trying to be condescending or judgmental or aggressive toward another person for not agreeing with your product choices I immediately feel suspicious. Can’t you see how these people act? Can’t you see how strange it is to defend a thing you pay money for that has terrible economics? If it wasn’t the “in” thing, being an AI person would be considered really weird. I look forward to the day it is. I hope you guys like having the stuff you said since 2022 repeated back to you! I’ve been saving it all. Time is running out for a graceful bow, and you better act quick!  If you feel self conscious while other people dunk on AI, that’s weird! I see people say they don’t like Macs all the time. Who gives a fuck! I’m not going to go to the mat for Tim Cook. People can make their own decisions.  Those comparing AI to AOL mailing CDs to people should feel ashamed of themselves. This is like if every single time you opened a magazine an AOL CD flew at your head, your boss told you he would replace you with a modem if you didn’t go online, and the news constantly ran segments called “I didn’t receive an email: father forgets son forever because he wasn’t online” or panels with “Internet experts” who said “I am on the Internet superhighway right now, and I’m certain that within 10 years AOL Time Warner will be able to email myself to my dad.”  Imagine if Shingy was a billionaire and went on TV every day in 1999 and told you “ the world must get ready, because you’re about to get a ICQ message from The Lord .” Generative AI was purpose-built to grift an economy run by executives and managers who don’t actually do any work. Its success has been driven by a remarkable, society-wide ignorance in the management sect, and its continued proliferation is only possible through the media’s continued trust and faith in the idea that CEOs are busy because they’re actually doing work. Yet even a Business Idiot eventually realizes that too much money is being spent, and the first one of these dimwits to cut their token budget will send the rest of them running for the doors. We should lock them. We should make everybody who obsessed over theoretical ideas about what AI can or will do ashamed for their intellectual deceit or constant ignorance.  At the end of the AI era, the only thing that will change the rot at the heart of our economy is the acceptance that the majority of companies are run by lazy, self-involved and ignorant fuckwits, and accountability for those who refused to scrutinize them. Microsoft has spent a total of $293.8 billion in capex since the beginning of Fiscal Year 2023 (which began in the back half of 2022). This means that around 30% of Microsoft’s capex ($87 billion) went to building OpenAI’s infrastructure. Based on discussions with sources familiar with Azure architecture, this is the vast majority of Microsoft’s operational capacity. AI revenues have to explode. Capex has to stop being invested. GPUs need to be margin positive, including both their cost and the debt associated with operationalizing them. AI revenue has to stay consistent both before and after you stop spending that capex. Microsoft’s RPOs jumped from $392 billion to $625 billion between Q1 and Q2 FY26 (or calendar year Q4 2025 and Q1 2026), driven by the $250 billion in “incremental Azure spend” from OpenAI (including already-existent commitments) locked up in October 2025 and the $30 billion promised as part of its deal with Anthropic from November 2025 . Based on Microsoft’s own disclosures , without Anthropic and OpenAI’s additions, RPO would have been effectively flat, as evidenced by the fact that in Q3FY26, remaining performance obligations sat at $627 billion .  Amazon’s RPOs jumped from $244 billion in Q4 2025 to $364 billion in Q1 2026, driven by its February 2026 $100 billion expansion of its $38 billion compute deal from November, and its extended partnership with Anthropic for 5GW of compute capacity unattached from any kind of dollar number.  Google’s RPOs jumped from $242.8 billion in Q4 2025 to $467.6 billion in Q1 2026, driven by ( per The Information ) $200 billion in committed spend on TPUs and compute from Anthropic, meaning that it has expanded its future revenues by an unremarkable $24.8 billion when you remove Anthropic’s spend, when RPOs had previously jumped $85 billion between Q3 and Q4, likely driven by its compute deal from October 2025 . It’s fair to assume a chunk of the remaining RPOs are from its deal to rent TPUs to Meta , announced in February 2026, which makes it likely that it accounts for the majority of the remaining $24.8 billion. Silicon will get cheaper. They’ll start selling services. They’re profitable on inference. It’s an example of a typical working day. At 8:30AM, the engineer notes that confirmation rates in Dallas dropped 3% overnight.  ‘Dallas inventory spiked; buyers went from 3 showings to 7. The agent shows the pattern: we're hitting the same buyer 7 times in 24 hours with "tour confirmed" pings. They're overwhelmed; they're muting us.’ The line before this says: “I don't open the codebase — I open the spec and eval dashboard.” Half an hour later, the engineer changes the spec, which is then tested against previous data, showing an improvement.  “The PM and I review diffs, check guardrails, approve.” Diffs are “differences” — essentially comparing two versions of the same document to see which lines have been changed.  The code is then rolled out.  At 11AM, the senior engineer mentors a junior engineer:  ‘A junior engineer's rescheduling agent is failing evals. I ask one question: "What happens if the buyer picks a slot the seller just blocked while the agent is negotiating?" We identify the race condition and add a constraint: "Always re-check availability at confirmation time." She updates the spec and evals. The agent passes.’ It is absolutely adorable they’re pretending that they’ll have junior engineers if this hellscape vision comes to life.  You can’t say “burn as many tokens as possible,” because employees will — as happened at both Amazon and Meta — deliberately create ways to burn more tokens using scripts and automations.  You can’t say “use AI every day,” because even if they do so, that doesn’t actually set up a success criteria. You can’t tell software engineers to try and “ship more software,” because that, again, emphasizes doing more, not making good stuff , and leads to an increase in velocity rather than how good the stuff is. You can’t say “pull requests” or any other metric a software engineer can manipulate, because in 100% of the situations where you give a software engineer a number to hit they will focus entirely on hitting that number.

Finance Shell AI Dart Business

▲ 0

0 views

A Working Library 2 months ago

Into the gap

It is right that the murder of many people be mourned and lamented. It is right that a victor in war be received with funeral ceremonies. Tzu & Le Guin, Tao Te Ching , page 38 H ow are we to prevent war? asks Virginia Woolf in the winter of 1937, as photos of the Spanish Civil War pile up on her desk, with their broken bodies and broken buildings, and Hitler and Mussolini gather forces to the east, and her own government’s war budget reaches new extremes. War, she asserts—and you will agree—is a horror, a terror that must be stopped. As well we know, confronted as we are with real-time video of genocide in Palestine, the massacre of school children in Iran, a fascist leader not abroad but in our own demolished house, asserting his right to make war wherever he likes, whenever he wants, including in our own cities, as armies under other names murder and disappear our neighbors with impunity. But, Woolf asks, what is she to do, what are the daughters of educated men to do in the face of that horror? And what are we, generations later, working women and their allies, how are we to stop it? It’s a good question, and we must spend some time trying to answer it. Woolf begins by considering how women might influence the decision to go to war and we may well begin with the same. To influence, we must have some knowledge to impart, some skill in speaking of it, and a listener who would hear us. We have some knowledge—the knowledge that war is a horror, the knowledge that when a missile falls from the sky and rends bodies into pieces that a terrible evil has been done. We can speak of this too, can point to the photos and videos that flit across our screens, children with missing limbs begging for food amid the ruins. These are images and reports of atrocity, undeniably and unequivocally. Yet who would listen, and how? Where can these words be spoken? Here we find we are in some trouble, for the supreme form of speech in our time is not words but money, both in legal doctrine and in fact of order, with our media controlled and manipulated by an obscenely wealthy few who have gobbled up platforms and papers and perverted them to their own aims, aims that seem very much in favor of war, for war has ever been the commander of wealth. When we speak against war we find our words drowned out, lost in the deepfakes and the advertising, the psyops and the slop, the stock market reports, the casual declarations of war crimes, the oil futures, the gilded festivities, the chattering and nattering among a purportedly progressive political class concerned with the appearance of civility but indifferent to its obligations. No knowledge moves through such mediums, only information, a ravening, unending stream of data in which knowing anything is nigh impossible. And such is that information that it is frequently as odious as the war it both directly and indirectly leads to: racism, misogyny, eugenics, transphobia. (That last a word that implies fear or aversion when the reality is much more violent, both speech and act that seek to eliminate a people whose courage in seeking their own liberty is among our brightest beacons.) But are these notions not the collaborators and soldiers of capital, and so of war? Are not racism and misogyny the masked recruits who go door to door, kitchen to bedroom to workplace, demanding labor and loyalty and love from an underclass who are threatened with suffering and death if they do not deliver it? Toni Morrison, whose words we may yet remember, said: “And they never, ever thought we were inhuman. You don’t give your children over to the care of people whom you believe to be inhuman….They were only, and simply, and now interested in the acquisition of wealth, and the status quo of the poor.” 1 Racism and eugenics were invented to justify the colonization of Black bodies just as sexism justified the enclosure of women’s. 2 The racists and misogynists of today work the same power: they create a world in which a few wealthy men dictate the material conditions of the lives of millions of others who must serve them, who toil for scraps, whose every step, however small, towards more freedom is violently and immediately resisted, and with overwhelming force—an impulse that you will agree is very much like the impulse to war. Look no further than the disproportionate attack on DEI, an effort that saw not to upend capitalism but merely to lightly expand the number of people who might not be entirely crushed by it, but which has been met with an extraordinary campaign to cancel huge swaths of scientific research, retract life-giving knowledge of medical care, hollow out our universities, purge career civil servants and leaders of the armed forces, and to eviscerate the federal workforce 3 —upending millions of lives and leaving our federal government, already poor from decades of neoliberal retreat, unable to deliver on the basic requirements for the life and liberty of its now abandoned public. That the federal workforce has long been one of the best chances for a comfortable life for Black and brown women excluded from comparable employment in the private sector is of course no coincidence. Meanwhile, the barons of the private sector have likewise backed down from even superficial concern for equality, and now demand such extreme fealty to their enterprises that only someone with no caretaking responsibilities whatsoever—with no care at all, not even for themselves—could possibly meet them. “Influence must be combined with wealth in order to be effective as a political weapon,” 4 Woolf concludes, and we grieve that the only change we can see in the century since is that the gap of wealth has widened, the effectiveness or lack thereof become only more extreme. Woolf was a member of the propertied class, but it was in her lifetime that women earned the right to their own property and were granted access to professional work, such that they might not be entirely in debt to their fathers and husbands. And yet in her time women secretaries were said to be routinely “fagged out” in the afternoons because they couldn’t afford a proper lunch. 5 Today, our food pantries work overtime to feed the working poor, people who work full time and more but don’t make enough to buy bread. Those who do make enough to live on do so in awareness of their intense precarity, the knowledge that they are one illness or storm away from ruin. And even the wealthiest worker has little compared to the investor class pushing for war, those who see war not as an abomination but as yet another opportunity to increase their bloated purse. What is our wealth compared to the billions spent on fighter jets, the $2.5 million spent on a single Tomahawk as it tears through a school full of little girls? What is our wealth compared to the mind-boggling quantities spent on the drones and satellites that make death as easy as clicking a button from the safety of a desk on the other side of the world? The same flick of a thumb can reduce a hospital to rubble or post a racist meme, often one right after the other. What is our wealth compared to the record-breaking $1.5 trillion requested for the military, a military that is already the richest on the planet ? Trump : “We have a virtually unlimited supply of these weapons. Wars can be fought ‘forever.’” So if money is influence, our relative influence has waned with the rise of the billionaire class. Woolf, recognizing the same, turns her attention instead to education. For if perhaps enough money cannot be mustered to prevent war, then learning—with its values of intellect and reason and enlightenment—may work in our favor, inasmuch as learning grows those faculties of reason, and reason is quite the antidote to the unreason of war. But again we find a problem. In Woolf’s time, while women have ostensibly been permitted into the colleges, they remain excluded from universities, and the women’s colleges are beggarly compared to those gleaming towers. Nor have women been permitted to adorn their names with the same letters and credentials that the men claim, a factor that keeps them from competing for the jobs that require them. It seems that the colleges are less places of learning than they are places of acquiring prestige, a prestige that is fiercely defended and protected, for prestige is a strangely fragile creature who can live only in scarcity and when exposed to too many of its own kind withers and dies like a tree choked by vines. And today? Well, women have torn down the gates to the universities, that much is clear. Women make up a majority of all college students in the US, and would be an even greater portion were it not for policies that directly work to balance the gender of student bodies . But that tearing down has been met by what can nearly be termed a war itself: a livid and indignant assault on places of learning from the men who want war, aiming at what has become the heart of the university, its beating and bloodied endowment. And the universities have, nearly to the letter, capitulated and retreated in the face of that assault, trading away centuries of purported intellectual freedom in order to protect the money needed to continue to operate, as if operating without that freedom was worth any money at all. Woolf writes: Is that not enough? Need we collect more facts from history and biography to prove our statement that all attempt to influence the young against war through education they receive at universities must be abandoned? For do they not prove that education, the finest education in the world, does not teach people to hate force, but to use it? Do they not prove that education, far from teaching the educated generosity and magnanimity, makes them on the contrary so anxious to keep their possessions, that “grandeur and power” of which the poet speaks, in their own hands, that they will use not force but much subtler methods than force when they are asked to share them? And are not force and possessiveness very closely connected with war? Woolf, Three Guineas , page 193 We see that same force and possessiveness in our own time: billions extorted from the universities, while the universities call in cops in riot gear —gear so named because when worn it inspires one to riot—to descend on students protesting genocide in Palestine. A great irony this would be, if irony were not the first casualty of war. For these brave students were met with war while exercising their right to protest the same, a right which past wars have been fought to defend but in which we seem to have retroactively declared defeat. Places of learning are always the first target of the fascist, because they are places that might counter the propaganda and pseudo-culture that leave us either pacified and accepting of their scraps or else fighting each other instead of fighting those who would start a war. Learning and thinking —a skill the billionaires are trying to supplant with machines that purport to think for us —are a challenge to the illogic and madness of war. To see an image of the broken bodies and broken buildings, to hear the testimony of those who lived, to have the skill and fortitude to ask how this could have happened, who benefits from such a horror, and how they might be stopped—for they must be stopped—is to exercise a lively mind and spirit, one capable of making the imaginative leap between the way things are and the way things ought to be. That interrogative and thinking mind is a threat to the fascist, who needs you to see things only as he does, who needs you unthinking and unquestioning, because only an unthinking and unquestioning mind could possibly accept the horrors of war. Only a mind so subdued by slop and propaganda and advertising, a mind unpracticed in observation and inquiry and imagination—only such a mind could be complacent as its pockets are picked to fund that most terrible of horrors. And so at last we turn to the workplace, as Woolf does, not in the hope that we might make enough money to counter the warmongers—for we have done the math, and no matter how hard we try, there is no chance of that—but because work is where we may, if we’re lucky, earn enough to keep a roof over our head and food in our belly, both of which are necessary to be able to think and act in the world. And we must be able to think, to remember that war is a horror, to resist being anesthetized by the memes and the vapid statements to violence. But here we find a curious contradiction: on the one hand, we are threatened with a lack of work , with our jobs taken over by machines who will never know that war is a horror, because they cannot know anything at all. On the other, high-pitched edicts that we must work so hard that there can be no time to think of anything else, no time to consider how these pictures of broken bodies and broken buildings came to be. ( Musk : workers “need to be ‘extremely hardcore,’ logging ‘long hours at high intensity.’”) How can both of these claims be true? How can the investor class simultaneously threaten us with no work, and, at the same time, threaten us with too much? It seems they fear equality more than hypocrisy. Perhaps we should also fear the disposition that the professions—which women fought so hard to enter, and now must fight so hard in which to stay—train us for. Here again is Woolf: And those opinions cause us to doubt and criticize and question the value of professional life—not its cash value; that is great; but its spiritual, its moral, its intellectual value. They make us of the opinion that if people are highly successful in their professions they lose their senses. Sight goes. They have no time to look at the pictures. Sound goes. They have no time to listen to music. Speech goes. They have no time for conversation. They lose their sense of proportion—the relations between one thing and another. Humanity goes. Money becomes so important they must work by night as well as by day. Health goes. And so competitive do they become that they will not share their work with others though they have more than they can do themselves. What then remains of a human being who has lost sight, and sound, and sense of proportion? Only a cripple in a cave. 6 Woolf, Three Guineas , page 258 It’s interesting to think with Woolf about our current march towards war, as the differences between her time and ours are revealing as much for what hasn’t changed. She wrote at a time when women were still largely excluded from professional work, from universities, from the armed forces. We read her today as women with one or more degrees, with careers, many of us carrying medals won in war zones and the scars to prove them, many of us with pips on our collar, credentials as long as those held by the men who guarded the libraries from the presence of women in Woolf’s time. But in both eras our presence in these places seems to have inspired an extraordinary, and extraordinarily violent, response. The assault against diversity programs is so out of proportion to those programs’ actual impact that we must admit something more elemental is going on: women’s presence in previously precluded spaces (and it is important to note that it is white women who have been the greatest benefactors of diversity initiatives, and Black and brown women who now suffer the greatest costs of their retreat) has inspired a level of violence among a small group of rich, insecure men that they will lay waste to the whole world before they will consider sharing their table with women as equals. Their own self-worth is so mean and spare that it withers when it comes into contact with those who do not bow and bend in their presence. The armed thugs marching through our streets, the speeches about force, force in our own cities, force elsewhere in the world, soldiers rechristened as “warfighters,” all of this is an assertion of manhood, a manhood reduced to nothing more than domination in all things, a masculinity that can see itself only in the violent oppression of others, whether that is other countries, other cultures, other races, other genders, or the more-than-human world. As Jamelle Bouie notes , “the vision of the world here is the vision of a rapist.” We are forced to conclude that to be in possession of a great deal of money, to be in a position of great authority, whether over an institution of learning or of government or of business, is to be in favor of war. The prestige and power that accompany both rank and great wealth—wealth which in our own day has grown so large as to be incomprehensible—also engender an instinct to possession and to the violent and disproportionate defense of that wealth. While we, who have neither great rank nor great wealth, know war to be an abomination, a horror through and through. Yet we can never hope to compete with the warmongers in either arms or cash, in prestige or status. So what are we to do? We must refuse to compete at all. We, with our empty hands , know it is right to mourn and lament the murder of many people. And so we mourn, and we lament, and we demand that our would-be leaders stop this incessant and evil warmaking. Are those demands enough? It would seem not. It would seem that despite great opposition to war , despite great risk to our economy, to our own safety as we shred our oldest and strongest alliances, that our demands for an end to war land on ears not deaf but blocked, stoppered with ego and greed and lust for domination in all its forms. And perhaps this should be no surprise. For why would a class of people so threatened by the mere presence of women in their schools and governments and workplaces ever open their ears to those women’s demands? Our speech must be a very great threat if they are so unwilling to hear it. So to speak against war is necessary—necessary for us to speak so with one another, so that we do not forget that war is a horror—yet insufficient. It is not enough to speak against war, for the warmongers, with their infinite money and infinite weapons, cannot hear us against the drums they so loudly bang for war. We must look elsewhere for the path that leads away from here. When Woolf was writing, women were precluded from the armed forces, and so could not refuse war by refusing to fight. We today are not subject to the same prohibition. We find ourselves among the ranks of soldiers both on our own soil and on many others. We have not earned the same respect, for many of our brothers seem to believe we have been put there solely for their use and abuse , and others—the same people who drive us to war, who claim no reason for war save war itself— work to exclude us once again . Yet women make up roughly a sixth of the armed forces , and perhaps as much of the forces in our streets. 7 Here is perhaps our greatest opportunity to halt the march to war. For we have it within our power to refuse to fight. We who know that war is a horror must refuse to raise a gun or fly a jet or steer a drone heavy with death into homes and hospitals and schools. We must refuse to go door to door in our own cities dragging people without warrant or reason into filthy, inhumane, and hastily built camps—for as sure as killing is a part of war, so too is gathering people up and locking them away. We must drop guns and kevlar and gas masks and walk away from the field of war, whether that field is distant from our homes or just down the street. We may look here to the courage of those like Ella Keidar Greenberg , an Israeli who, at 16 years of age, signed a pledge refusing to enlist in the military and was then, at 19, jailed for that refusal. “Refusal is the imperative,” she speaks, and we who have not plugged up our ears to reason and wisdom can yet hear her, and agree. For to make the horror of war with your own hands is to become a horror yourself. 8 This is easy to say for the great many of us who do not fight in war, who have not raised guns or donned armor or placed hands on keyboards and rained death on schools and hospitals from afar. But the imperative to refusal remains: we must refuse to lend our hands or minds to war, in whatever way we can. And so we must also refuse to work for war, to use our labor to make the technology of war, whether of weapons or of surveillance or of detention, whether that technology is used in our own streets or somewhere afar—for any technology used afar will come home soon enough, as we see with the militaries in our streets, outfit with cast offs from so many wars abroad. 9 We must not lend our hand to the making of guns or missiles or drones, of targeting systems or intelligence databases, of satellites that scour the planet for schools and hospitals, of algorithms that prescribe processes for murder, processes that promise to scrub their operators clean of the blood that follows but which will haunt them, nonetheless. Is this enough? It is not. For war is such an enormous undertaking—witness the trillions of dollars, an amount of money too big to think with—that it seeps into nearly every part of the economy. The same servers that summon servants to your door are used to surveil the people of Gaza; the same newspaper that brings details of the war to our eyes and ears also perpetuates a story that the greatest hardship of war is the price of gas at the pump. The same so-called AI that makes it easier to prototype a website is simultaneously being used to generate enormous quantities of racist and misogynist slop that treats war like a spectator sport. The same university that teaches the history of war also pays millions in bribes to the warmongers, while making a concerted effort to erase trans people from the very same history books. If we are truly committed to not working for war, we must not work for any of it. Not for the weapons manufacturers or the drone makers or the algorithm authors; not for the papers or the products or the schools. Perhaps you will think I am being too harsh. Perhaps you will say, but this is my only way of making a living, of keeping a roof over my head and my children’s heads, of feeding and clothing my loved ones. After all, we have also noted how our publics have been decimated by the very same men who push for war, men who have likewise colluded to raise prices on milk and eggs, who have transformed homes into commodities, such that we who had so little money compared to them seem every day to have less and less. Already our food pantries work overtime feeding the working poor, and we rightly fear every cough and tooth ache, every flutter of our overworked hearts or tiny lump beneath our skin, for medicine is increasingly a privilege reserved only for the rich. How could we refuse work under such conditions, when work is increasingly scarce? Here we must pause and again wonder at that scarcity. For it is a curious thing that work is becoming harder and harder to come by, that what work there is is often so poorly remunerated we must visit the pantries for bread at the end of the workday. Or, if it pays well, it does so under the constant threat that it could end at any moment, that it will end soon enough. Is it not the case that the men who loudly bang the drums for war, who build the technologies of surveillance that are used both to round people up and to aim missiles on their backs, who pollute our skies with satellites and insert themselves into the field of war as if they were heads of state themselves, states of ego and greed and impunity—are these not the selfsame men who declare we no longer need workers at all, that one machine can do the work of dozens? And do they not declare, out of the very same mouths, with the very same breaths, that those few workers who remain must work themselves to the bone, must work every waking hour they can, must eschew rest and play and leisure for the work is too great to put down for even a moment? And do they not also say—for as we have seen, those with more money have more speech, and seem ever to want us to hear them—that it is immigrants who are taking away all the jobs ? (A dog-ate-my-homework excuse, if there ever was one.) And meanwhile there is so much work that needs doing but isn’t being done: our schools overcrowded, our farms short-handed, our streets and bridges crumbling, our parks neglected, our clinics overrun, our laboratories empty. This is not to say that the scarcity isn’t real. It is real enough, as the lines at the food pantries attest. But it is manufactured ; it is built bolt by chip by screw by a billionaire class who want workers who complain neither of their warmongering nor of their whip. On the one hand, they threaten us with no work at all, with the misery and penury that comes from a lack of work, and therefore a lack of the means of living. On the other, they demand endless work, a work that wipes out all other avenues for thinking and being, that leaves us programmable and programmed, no space left in our minds for thoughts they haven’t placed there. Are we to merely acquiesce, to accept their scraps and the miserable conditions attached to them? Surely not. For if we accept these conditions, will they not impose even worse upon us? Will they not keep increasing their demands and decreasing our pay until we are working ceaselessly, and for nothing? What would compel them to stop? Already we have seen that their greed for money and for power is so voracious it will tear through buildings and through bodies, it will murder many people, it will poison the air and the soil, it will bring great storms upon us. So there must be an end, and it is only we who can bring that end about. So I say again we must refuse to work for war. But I do not wish you any hardship. If the only work available to you is the work of war, or work that has been perverted to the aim of war—and I am trusting that you have done your best to find other work, to make your living in a manner that does not end the lives of others—then there remain yet other avenues to take. Here you must gather with your colleagues and comrades, for the work against war is not solitary. You must first speak and be heard by each other, know that you are not alone in recognizing that war is an abomination, a great and terrible horror. For while speaking into the networks and the platforms is like speaking to the wind, your words tossed away from you before they can reach your own ears, we still have the ability to speak to our colleagues and to our neighbors, to speak unmediated and uncensored with each other. To speak with our mouths and with our hearts and with our lively, imaginative minds. To say, war is a horror, and I will not work for it, and are you with me? Can we speak together? Can we move and act against war hand in hand, and right here, where we stand? Here we see a great many of our kith and kin already stepping up. We can look to workers at Amazon , Google , Salesforce , and others who demand that their work not be used for surveillance, mass deportation, drone warfare, or genocide. We can look to the hundreds of workers at Thomson Reuters who raised alarms after learning that their company was selling data to ICE, prompting shareholders to demand an investigation . We can look to the community in Monterey Park, California , who successfully organized in favor of a ban on the construction of data centers—after noting that in addition to being polluting, noisy, energy guzzlers, such data centers also fuel ICE’s violence against their own neighbors. We can look to the Harvard graduate students currently on strike, whose demands include protections for international students at risk of deportation. We can look to the twenty-four attorneys general who have filed more than seventy lawsuits aimed at stopping the administration from waging war at home. And we can look to Luanne James, a librarian in Tennessee, who when asked to remove books from her library—books flagged for such transgressions as “female empowerment” and “following one’s dreams”—said, “ I will not comply. ” For is not censorship likewise a tool of war? Haven’t the book burners and the warmongers always been the same people, with the same aim? Are not slop and chatbots who care nothing for veracity the new tools for censorship—censorship by means of pollution rather than prohibition, but the ends are the same. James was subsequently fired for her dissent. 10 Refusal always invites consequences. But then so too does compliance, and often very grave consequences at that. Here we may heed the advice of the veteran scientists who resigned from the National Institutes of Health after it was gutted by the Trump administration. They implore , “Please decide where your red line is so you can choose to act before the line is already behind you.” There is risk here, of course. Organizing is, in theory at least, a protected activity and legally you may not be retaliated for it, but we have seen who the law protects and who it bends and breaks for and have no confidence in it protecting the likes of us. But there is risk no matter what we do or do not do. To be alive, to have a body vulnerable to gun and missile and chemical weapon, to famine and to thirst, to penury and hardship, is to be at risk; only the dead are relieved of the risk of harm. Your employer may punish you for organizing, but what is that risk compared to the risk of being complicit in war? The risk of knowing yourself to be someone who helped rain death on schoolchildren, who helped imprison your fellow workers in filthy detention camps, who helped program people’s minds to be numb to atrocity and horror? For you will know what you have done. Even if your daytime self can wrap you up in comforting excuses and justifications, can be lulled by the distractions and the advertisements and the television that anesthetizes your conscience, you will know it in the dark of the night. Our dreams know where we have gone wrong and they will never let us forget it. 11 But perhaps even this risk seems too great. You know your circumstances, and you know the ways the investor class has of keeping your head down. You cannot be fairly asked to put your own life, or your kin’s lives, on the line. And yet you are not without the ability to work against war, even in these difficult times. For you can work against war while seeming to work for it. Perform your work diplomatically while leaking information to the press, so that those on the outside who are safe from retaliation may organize in your stead. 12 Look for ways to gum up the works; raise concerns and questions and show where plans are short, where steps have not been thought out, where coordination is insufficient. Do not meet expectations but dash them, show them to be shortsighted or foolhardy, lacking sufficient detail; make those who set them doubt their own understanding of the world (as they try to sow doubt in you). They have made this easy on you, the warmongers and profiteers, by foisting unpredictable and inconstant machines upon you and mandating their use, by setting irrational milestones that could never have been met even by those who tried. Right there is a ready-made excuse for why the work could not be delivered as asked—your hands were tied. Do the work if you must, but do it dragging your feet, do it always on the lookout for ways to slow down the march to war and so give others the time to stop it. Does this gall you? It galls me. We ought not to have to spend our energy, what little and precious time we have on this earth, denigrating and diminishing our own skills. It is a violence to the self to do our work poorly. But against the alternative—against setting those same skills in the making of war—it seems a small sacrifice, and a necessary one. For it is not only your skill in, say, design or management or engineering that you may exercise. It is also the skill of refusal, the skill of refraining from making war in all its many and terrible forms. And that too is a kind of work, a good work, work that all of us can do. For there is one weapon that only we possess and which the billionaires and the warmongers can never take from us. One weapon which so frightens them they will twist their words into knots, they will spend the entirety of their vast fortunes trying and failing to convince us that we don’t possess it at all, they will claim over and over and without evidence that it is vanishing before our eyes even as it remains right there in our hands, clear and plain to hearts yet open to the world: the refusal to work. To refuse is a creative act. What is created in a refusal is a gap, a space, a moment in which something else makes ready to emerge, something that waits upon our invitation and a bit of water or sunlight to pop itself out and set down roots. To refuse is to create that which can only exist in the shade of that refusal, the refusal giving shelter to the choice that appears behind it. To refuse is to choose. In that choice, we find ourselves in the gap, in the place where no one has programmed our thinking, no one has told us what to do, no one has left any instructions or orders that we must follow. No one stands ready to answer our questions or to assign us tasks or to relieve the anxiety of being alive to uncertainty, for this has always and ever been the only way to be alive. In this gap is not one choice but many, a myriad of choices, for from here on out there can be no prescription, no map or plan or diagram. Only one step, and then the next. Yet we are not without skill or art. In fact, it is our art which is most at need here, our art that helps us imagine how things could be different, how we could work not for war but for peace, and for liberty, and for care for all our kin in all the kingdoms. How we could live with one another if prestige and missiles and extreme wealth were relegated to the history books, where they belong. It is our art, the art of painting or drawing or sculpting or dancing or making music or writing—and while all the arts are needed here, I will make a special plea for writing as that which so often gives us new worlds to think with—that we can think with the question of what we are to make with one another when we refuse to make war. For to refuse the work of war is to choose to see things as they really are, and as they yet could be. This is a choice we make most strongly when we make our art, when we bring our keen attention to the world and do not flinch from it, do not numb ourselves to it, but rather look at it squarely and know that however things are, they can—they will —be otherwise. What could our work become when it isn’t the work of death, of domination, of separation and detention and surveillance? What is our work when we give up seeking wealth and prestige—which no matter how hard we work, we can never have enough of? What is our work when we do not accede to orders from above but make choices with each other? What is our work when we see it not as a way to make a wage but a way to make more life , not only for ourselves, but for everyone? What becomes of our work if we work for the living? To refuse is an ending; an ending to our work being used to rend buildings and bodies, to massacre schoolchildren, to surveil and capture and detain. To refuse is a beginning. To turn away from the work of war is to turn toward the work of making a living world, work that does not answer to the billionaires, with their slavering, unending greed, but which only answers to each other. The gap that we create with our refusal is not void but potential, not emptiness in the sense of want but empty as a bowl or bag is empty, as an ear cocked to a speaker, a pair of hands cupped and raised to the roiling and darkening sky. From A Humanist View , a speech given at Portland State University in 1975. Quoted in Táíwò, Reconsidering Reparations , page 6. Táíwò adds, astutely, “Racism was only ever a smoke screen.”  ↩︎ “[I]n pre-capitalist Europe, women’s subordination to men had been tempered by the fact that they had access to the commons and other communal assets, while in the new capitalist regime, women themselves became the commons, as their work was defined as a natural resource, laying outside the sphere of market relations.” Federici, Caliban and the Witch , page 97  ↩︎ For just some examples of these efforts, see Unbreaking’s explanations of the assaults on the federal workforce , medical research funding , and trans healthcare .  ↩︎ Three Guineas , page 170.  ↩︎ Ibid, page 404.  ↩︎ This is clearly a reference to Plato’s cave, and the comparison hits a little harder in our own time: the shadows on the cave wall have been compressed to the mirrored screens we hold in our hands.  ↩︎ A since-deleted page on the ICE website says that women made up 15% of law enforcement officers employed by ICE as of 2023 ( archive link ). That the page has been deleted perhaps says something about how little ICE cares for the women in its employ.  ↩︎ The Center on Conscience and War reports that it has seen a 1,000% increase in US service members interested in becoming conscientious objectors since the start of the Iran war. Mike Prysner, the Center’s director says, “I haven’t heard from a single caller who said, ‘I’m scared of dying in a war I don’t believe in.’ All of them are scared of killing people in a war they don’t believe in.”  ↩︎ Aimé Césaire termed this the “boomerang” effect .  ↩︎ A legal defense fund has been set up to help James contest her termination.  ↩︎ In The Third Reich of Dreams , Beradt reports that those who worked against the Nazis had dreams of fierce hope, while those who collaborated and capitulated were wrought by nightmares of terror and humiliation.  ↩︎ The Freedom of the Press Foundation maintains some good advice on how to protect yourself while sharing information with the press—including the counsel to avoid visiting this link from a device your employer controls.  ↩︎ View this post on the web , subscribe to the newsletter , or reply via email . From A Humanist View , a speech given at Portland State University in 1975. Quoted in Táíwò, Reconsidering Reparations , page 6. Táíwò adds, astutely, “Racism was only ever a smoke screen.”  ↩︎ “[I]n pre-capitalist Europe, women’s subordination to men had been tempered by the fact that they had access to the commons and other communal assets, while in the new capitalist regime, women themselves became the commons, as their work was defined as a natural resource, laying outside the sphere of market relations.” Federici, Caliban and the Witch , page 97  ↩︎ For just some examples of these efforts, see Unbreaking’s explanations of the assaults on the federal workforce , medical research funding , and trans healthcare .  ↩︎ Three Guineas , page 170.  ↩︎ Ibid, page 404.  ↩︎ This is clearly a reference to Plato’s cave, and the comparison hits a little harder in our own time: the shadows on the cave wall have been compressed to the mirrored screens we hold in our hands.  ↩︎ A since-deleted page on the ICE website says that women made up 15% of law enforcement officers employed by ICE as of 2023 ( archive link ). That the page has been deleted perhaps says something about how little ICE cares for the women in its employ.  ↩︎ The Center on Conscience and War reports that it has seen a 1,000% increase in US service members interested in becoming conscientious objectors since the start of the Iran war. Mike Prysner, the Center’s director says, “I haven’t heard from a single caller who said, ‘I’m scared of dying in a war I don’t believe in.’ All of them are scared of killing people in a war they don’t believe in.”  ↩︎ Aimé Césaire termed this the “boomerang” effect .  ↩︎ A legal defense fund has been set up to help James contest her termination.  ↩︎ In The Third Reich of Dreams , Beradt reports that those who worked against the Nazis had dreams of fierce hope, while those who collaborated and capitulated were wrought by nightmares of terror and humiliation.  ↩︎ The Freedom of the Press Foundation maintains some good advice on how to protect yourself while sharing information with the press—including the counsel to avoid visiting this link from a device your employer controls.  ↩︎

Culture Writing Dart

▲ 0

0 views

Hugo 4 months ago

Dogfooding: Why I Migrated My Own Blog to Writizzy

In 2022, I created an open-source static blog generator: Bloggrify . It’s conceptually similar to Hugo —it generates a static site (just a bunch of HTML files) that you can host for free on Cloudflare, GitHub Pages, or Bunny.net . Before that, I had tried everything: WordPress, Joomla, Medium. I wanted to regain flexibility and customize my blog exactly how I wanted. But let’s be honest: I’m a developer, and I mainly wanted a new technical playground. Fast forward to 2026, and I have to admit: using a static blog has become a major friction point for my writing. So, I decided to migrate again, this time to a managed platform: Writizzy , another product I’m building. This move is a great opportunity to talk about several things: Bloggrify started as a love letter to the Nuxt ecosystem, specifically . Back when I migrated from WordPress, my criteria were simple: In 2022, it wasn't a "product" yet—just my personal blog code made public. It only became a full-fledged open-source project in 2024, with a dedicated site and a proper README to encourage contributions. I wanted the product to be "opinionated." Nuxt-content does 90% of the heavy lifting, but it’s a generic tool. For a real blog, you still need to build the RSS feed, sitemap, robots.txt, comments, table of contents, share buttons, newsletter integration, analytics, and SEO. That’s what Bloggrify is: a "starter pack" that comes with everything pre-configured. Think of it as Docus , but for blogs instead of documentation. I’m a numbers person. When I launch a project, I want to see usage. It might sound trivial, but considering the effort it takes to manage NPM releases (which is honestly a nightmare), handle versioning, and maintain themes, you expect a minimum return on investment. Bloggrify reached 164 stars on GitHub and sits somewhere in the middle of the pack on Jamstack.org . That’s... okay, I guess. But in reality, I have almost zero feedback on its actual usage. A few rare GitHub issues, one contributor who was active for a few weeks before vanishing, and then silence. I only know of one blog that used it before switching back to Hugo. The experience has been bittersweet. Building in the dark is demotivating. However, it did lead me to launch two other side-products: I launched Broadcast and Pulse in 2024 and 2025. They’re living a quiet life, but they aren't "exploding." My target audience is static bloggers—mostly developers. And as we know, developers are the hardest group to convince to pay for a service! Still, I’m satisfied. These products taught me how to build a SaaS, handle subscriptions, and find my ideal tech stack. My own newsletters were sent via Broadcast (reaching about 150 subscribers), and I used Pulse to track which articles were actually being read. The reality? These two tools generate about €100 in Monthly Recurring Revenue (MRR) . Not enough to retire on, but a great learning experience. And that brings us to Writizzy. With Bloggrify, I realized my writing workflow had become painful. Between maintaining the framework, jumping between spell-checkers, writing in Markdown, spinning up a local server to check for broken links, and waiting for build/deployment times... I was losing hours. For my last article, someone pointed out a few typos. It took me 20 minutes between editing the file and seeing the fix live. Add to that the friction of managing images in an IDE and the recent Nuxt 4 / Nuxt-content updates which, while I love them, have made the dev experience slightly slower for simple blogging. To be honest, I wasn't aware of that. I put up with these inconveniences and was still very happy to have “flexibility” in what I could do with my blog. I wasn't fully aware of this "friction" until I built Writizzy . Writizzy is the synthesis of my blogging experience. It’s a mix of Substack, Ghost, and Medium, but built as a European alternative with four core pillars: I moved my English blog to Writizzy first (this one), with no intention of moving the french one. But I soon noticed I was writing much faster on the English site. The workflow was just... better. Copy-pasting images directly into the editor, instant previews, no server to launch. It was a joy. I hesitated for a long time before migrating eventuallycoding.com . I knew that by doing so, I was taking the risk of killing Bloggrify. If even I don't use it anymore, the project enters a danger zone. When you don’t use your own product daily—when you’re no longer obsessed with the problem it solves—it’s almost impossible to stay attached to it. This is a symptom I see in so many "disposable" projects across the internet: built by people who flutter from one idea to the next without any real skin in the game. So yes, moving away from Bloggrify is a risk. But I’ve come to terms with it. Today, I have almost zero evidence that Bloggrify is being used. Meanwhile, Writizzy already has 314 blogs and 11 paying users (€135 MRR) in just four months. Why stubbornly cling to Bloggrify? Ultimately, I believe I’m solving the same problem with Writizzy, but in a much better way. I receive feedback emails and feature requests every single week. I get constant positive reinforcement from people actually using it. The product isn’t perfect, but it improves every day. It improves because real users are pushing me to refine the site, fix what’s broken, and add the features that absolutely need to be there. And it also improves because I use it constantly. This is the massive benefit of dogfooding . Every day, I am confronted with my own software, so I know exactly what needs to change. So yes, Bloggrify is moving to maintenance mode. I’m taking this opportunity to turn all templates into Open Source. Two of them were "premium," but it wouldn't make sense to keep them that way today. I tell myself I’ll still evolve it from time to time, but honestly, I wonder if I’m just lying to myself. As for Hakanai.io , I’m definitely continuing. The problem it solves still fascinates me. I get great feedback, especially on Broadcast. Pulse , however, suffers from being misunderstood. It’s a "blog analytics" product, and people don't really grasp what that entails—SEO advice, outlier detection, evergreen content tracking. I’m not great at marketing, so it mostly flies under the radar except for the readers of this blog who took the time to test it. But I’m motivated to keep them alive. As for Writizzy , there is no doubt. The product is incredibly exciting to build. The stakes are high: building a platform for expression that exists outside the US-centric giants. The traction is there, and the numbers follow (+45% MoM user growth). Welcome to this blog, now officially on Writizzy. As a reader, you can already test several things: The Discover feed to read other articles from Writizzy bloggers. We’ve handpicked a few to start with, and this feed will become even more customizable in the future. Welcome home. Dogfooding: Why you absolutely must use your own products. The harsh reality of Open Source: Why it’s harder than it looks. Product Satisfaction: The joy of building something people actually use. The future of my projects: Bloggrify, Writizzy, and Hakanai.io . A simple templating language (Markdown). Extensibility (RSS feeds, sitemaps, etc.). Low carbon footprint (static sites are incredibly efficient). broadcast.hakanai.io : A newsletter system for static blogs based on RSS feeds. pulse.hakanai.io : A specialized analytics tool for bloggers (not just generic web traffic). Sustainability : Focusing on reversibility and interoperability. Discoverability. Economic accessibility : Implementing Purchasing Power Parity (PPP). Transparency. The comments section . The newsletter subscription (if you haven’t already).

Dart Open Source C# HTML Web Development

▲ 0

0 views

Corrode 5 months ago

Gama Space

Space exploration demands software that is reliable, efficient, and able to operate in the harshest environments imaginable. When a spacecraft deploys a solar sail millions of kilometers from Earth, there’s no room for memory bugs, race conditions, or software failures. This is where Rust’s robustness guarantees become mission-critical. In this episode, we speak with Sebastian Scholz, an engineer at Gama Space, a French company pioneering solar sail and drag sail technology for spacecraft propulsion and deorbiting. We explore how Rust is being used in aerospace applications, the unique challenges of developing software for space systems, and what it takes to build reliable embedded systems that operate beyond Earth’s atmosphere. CodeCrafters helps you become proficient in Rust by building real-world, production-grade projects. Learn hands-on by creating your own shell, HTTP server, Redis, Kafka, Git, SQLite, or DNS service from scratch. Start for free today and enjoy 40% off any paid plan by using this link . Gama Space is a French aerospace company founded in 2020 and headquartered in Ivry-sur-Seine, France. The company develops space propulsion and orbital technologies with a mission to keep space accessible. Their two main product lines are solar sails for deep space exploration using the sun’s infinite energy, and drag sails—the most effective way to deorbit satellites and combat space debris. After just two years of R&D, Gama successfully launched their satellite on a SpaceX Falcon 9. The Gama Alpha mission is a 6U cubesat weighing just 11 kilograms that deploys a large 73.3m² sail. With 48 employees, Gama is at the forefront of making space exploration more sustainable and accessible. Sebastian Scholz is an engineer at Gama Space, where he works on developing software systems for spacecraft propulsion technology. His work involves building reliable, safety-critical embedded systems that must operate flawlessly in the extreme conditions of space. Sebastian brings expertise in systems programming and embedded development to one of the most demanding environments for software engineering. GAMA-ALPHA - The demonstration satellite launched in January 2023 Ada - Safety-focused programming language used in aerospace probe-rs - Embedded debugging toolkit for Rust hyper - Fast and correct HTTP implementation for Rust Flutter - Google’s UI toolkit for cross-platform development UART - Very common low level communication protocol Hamming Codes - Error correction used to correct bit flips Rexus/Bexus - European project for sub-orbital experiments by students Embassy - The EMBedded ASsYnchronous framework CSP - The Cubesat Space Protocol std::num::NonZero - A number in Rust that can’t be 0 std::ffi::CString - A null-byte terminated String Rust in Production: KSAT - Our episode with Vegard about using Rust for Ground Station operations Rust in Production: Oxide - Our episode with Steve, mentioning Hubris Hubris - Oxide’s embedded operating system ZeroCopy - Transmute data in-place without allocations std::mem::transmute - Unsafe function to treat a memory section as a different type than before Gama Space Website Gama Space on LinkedIn Gama Space on Crunchbase

Programming Hardware Rust Dart Science

▲ 0

0 views

Michael Lynch 6 months ago

Refactoring English: Month 13

Hi, I’m Michael. I’m a software developer and founder of small, indie tech businesses. I’m currently working on a book called Refactoring English: Effective Writing for Software Developers . Every month, I publish a retrospective like this one to share how things are going with my book and my professional life overall. At the start of each month, I declare what I’d like to accomplish. Here’s how I did against those goals: The blog post was a risky bet because it only could reach new readers if it hit the front page of Hacker News, and its only chance of that is the first couple weeks of 2026. Fortunately, the post reached #1 on Hacker News and remained on the front page for almost 22 hours. It continues my strategy of highlighting other successful tech writers , a strategy I like because it feels like a win-win for me, readers, and the writers I showcase. I still have the Hacker News prediction game at about 80% complete. I’m not sure what to do with it because it’s almost done, but I feel like it’s not fun, so I’m never motivated to complete it. But I want to get it over the finish line to see what people think. Ironically, the chapters I’m working on are about motivation and focus, but I keep letting my experiments with MeshCore interfere with my writing. I’ve been better at maintaining focus in the new year, and distractions are actually helpful because I’m getting fresh experience to write about regaining focus. Again, I got distracted by MeshCore experiments in December and didn’t make as much progress as I wanted. I love design docs and find them helpful but they’re also incredibly boring to write, so it was always tempting to shelve the design doc for something with more instant gratification. Pre-sales are down because I didn’t have any new posts to attract new readers (I didn’t publish the Hacker News post until January). Still, it’s a positive sign that my “passive sales” continue to grow. In December, I had almost $500 in pre-sales. If I compare that to months with similar website visitors, May had only $241 in pre-sales, and August had $361, so the numbers are trending up. I hope that as the book grows more complete and more readers recommend it, the passive sales continue to rise without relying on me finding a successful marketing push each month. When I ran my Black Friday promotion in November, a reader emailed to say that 30% off (US$20) is still an unaffordable price in Argentina for a book. He asked if I’d consider regional pricing. He mentioned that Steam games are typically priced 50% lower in Argentina than the US, so I figured that was a good anchor. I collect payments through Stripe, and I couldn’t find any option for regional pricing in my Stripe dashboard. I found an article in Stripe’s knowledge base called “Geographic pricing in practice: Why it matters and how to implement it.” I was delighted until I read the entire article and discovered they’d forgotten to write the “and how to implement it” part. So, Stripe advocates for regional pricing, but they don’t actually offer it as an option. It was a helpful reminder that Stripe is the worst payment processor except for all those other payment processors . So, for my Argentinian customer, I used a one-off process where I manually created a custom payment link for him at a discounted price. And when I went through the process, I realized I could set the price in Argentine pesos so he wouldn’t have to pay a currency conversion fee. I set the price to 22,000 ARS (about US$15), and he seemed happy with the price and the checkout experience. The reader suggested I publicly offer regional pricing, at least for countries like Brazil and India, which have high numbers of developers but relatively low purchasing power. Even without native Stripe support for regional pricing, it seemed like it wouldn’t be that hard to automate the thing I did manually. I read about Sebastien Castiel implementing regional pricing for his course, which led me to Wes Bos’ post about the same thing . Sebastien shared a lot of technical details, but his solution was heavy on React, whereas my site is vanilla HTML and JavaScript. He also relied on discount codes, which I don’t like because it means most customers see that there’s a special deal they’re not getting. I spent a few hours implementing a solution using a cloud function that determines the right price on the fly and dynamically creates a Stripe checkout link. Then, I realized I could precompute everything and eliminate the need for server-side logic, so I deleted my cloud function. My implementation looks like this: The user just picks their country and it activates the Stripe purchase link for that country, and they pay in their own currency. I’m going by the honor system, so I don’t bother with IP geolocation or VPN prevention. I do hide the discount for each country to discourage people from picking the cheapest option. And part of the benefit of pricing in each country’s local currency is that if someone cheats and picks a region that’s not really their home currency, they lose some money in conversion fees. The numbers feel not quite correct. According to strict PPP, the equivalent of $30 in the US is $4 in Egypt, but I suspect you can’t really buy non-bootleg books for programmers in Egypt for $4. When Wes Bos did this, he just asked his readers to tell him fair prices, so I’ll try that too. Leave a comment or email me the normal price range for developer-oriented books in your country. In December, I published “My First Impressions of MeshCore Off-Grid Messaging.” I was excited about the technology but disappointed to discover that the clients are all closed-source . At that point, I decided to pause my exploration of MeshCore, but Frieder Schrempf , a MeshCore contributor, replied to my post with this interesting perspective : I share a lot of your thoughts on this topic. Personally I see the value of MeshCore in the protocol and not so much in the software implementations of the firmware, apps, etc. […] If MeshCore as a protocol succeeds and gets widely used (currently it looks like it does) then properly maintained open-source implementations will follow (at least I hope). I agreed with Frieder and thought, “Maybe I should just write a proof of concept open-source MeshCore app?” Actually, there already was a proof of concept MeshCore app. Liam Cottle, the developer of the official MeshCore app, previously wrote a web app for MeshCore as a prototype for the official version. He deprecated it when he made the official (proprietary) MeshCore app, but the source code for his prototype was still available, and the prototype had most of the features I needed. I wondered how difficult it would be to port the prototype to mobile. MeshCore is too hard to use as a web app, as it requires Bluetooth access and offline mode. I’ve heard somewhat positive things about Flutter , Google’s solution for cross-platform mobile development. I suspected that an LLM could successfully port the code from the web prototype to Flutter without much intervention from me. My plan was to have an LLM create a Flutter port of the prototype in three stages: That worked, but every step was clunkier than I anticipated: I thought it would be a quick weekend project I could whip together in a few hours. 30 hours and $200 in LLM credits later, I finally got it working. Running my MeshCore Flutter app on a real Android device But the day I got my Flutter implementation to feature parity with the prototype, I went to share it on Reddit and saw someone had just shared meshcore-open , a MeshCore client implementation in Flutter. It was the same idea I had but with far better execution. I was disappointed someone beat me to the punch, but I was also relieved. From my brief experience working with Flutter, I was eager to get away from Flutter as quickly as possible. I only wanted to make a proof of concept hoping someone else would pick it up, so I’m happy that there’s now an open-source, feature-rich MeshCore client implementation. While working on my MeshCore Flutter app, I had to implement low-level logic to parse MeshCore device-to-client messages. There’s a public spec that defines MeshCore’s peer-to-peer protocol, and even that’s fairly loose. But there’s another undocumented protocol for how a device running MeshCore firmware communicates with a companion client (e.g., an Android app) over Bluetooth or USB. The de facto reference implementation is the MeshCore firmware , but it intermingles peer-to-peer protocol logic with device-to-client protocol logic and UI logic, and it spreads the implementation across disparate places in the codebase. For example, a MeshCore client can fetch a list of contacts from a MeshCore device over Bluetooth, but it has to deserialize the raw bytes back into contacts. There’s no library for decoding the message, so each MeshCore client and library is rolling their own separate implementation: What I notice about those implementations: My first thought was to rewrite the logic using a protocol library like protobuf or Cap’n Proto , but I don’t see a backwards-compatible way of integrating a third-party library at this point. So, what if I wrote a core implementation of the MeshCore device-to-client protocol in C? I could add language-specific bindings so that we don’t need whole separate implementations for Dart, Python, JavaScript, and any other language you’d want to write in. So, I started my own MeshCore client library: The library is not ready to demo as a proof of concept, but it’s close. It’s entirely possible the MeshCore maintainers won’t like this idea, and it’s basically dead in the water without their buy-in. But I did it anyway because I’d never tried writing a cross-language library, and that was an interesting experience. The last time I tried to call C code from Python was 20 years ago, and I had to use SWIG . Back then, it felt painful and hacky, and it seems to have gotten 80% better. I desperately wanted the core implementation to be Zig rather than C, but I saw too many blockers: Is $30 (USD) for a developer-oriented book expensive where you live? If so, let me know what you’d expect to pay for a programming book like Designing Data-Intensive Applications in your country (in local currency). I added regional pricing for my book based on purchasing power parity. I created my first Flutter app. I’m writing my first cross-language library. Result : Published “The Most Popular Blogs of Hacker News in 2025” instead Result : Made progress on two chapters but didn’t complete them Result : Got 80% through a design doc draft Manually get a list of all countries / currencies that Stripe supports. Write a script that pulls data from the World Bank to calculate the purchasing power parity (PPP) for each country in the list. Calculate each country’s discount based on their purchasing power relative to the US. e.g., the PPP of Brazil is 54% lower than the US, so they get a 54% discount. Filter out countries where the PPP is within 15% of the US (too small a discount to bother). Filter out countries where the discount would be negative. Otherwise, customers in Luxembourg would have to pay double . Limit the discount to a maximum of 75% Otherwise the price in Egypt would be US$4, meaning I’d get like $3.50 after conversion fees. Automatically generate country-specific Stripe price objects and Stripe payment links for each country remaining in the list. Put all the countries in an HTML dropdown on my site: Write end-to-end tests for the prototype web app using Playwright. Port the prototype implementation to a Flutter web app, keeping the end-to-end tests constant to ensure feature parity. Add an Android build to the Flutter project. Before I could write end-to-end tests for the prototype, I had to convert it to use semantic HTML and ARIA attributes because a lot of the input labels were just bare s. I couldn’t keep the Playwright tests constant because Flutter actually doesn’t emit semantic HTML for web apps. It creates its own Flutter-specific HTML dialect and draws everything on an HTML canvas. Most Playwright element locators still work somehow, but I had to make a lot of Flutter-specific changes to the tests. It took a long time, even with an LLM, to figure out how to build an Android package with Flutter. Gradle, Android’s build system, is buggy on NixOS. I kept running into situations where it was failing with mysterious errors that eventually turned out to be stale data it had cached in my home directory. Flutter makes it surprisingly difficult to communicate over Bluetooth. On the web (at least on Chrome), you essentially get it for free by calling , but with Flutter, you have to use a proprietary third-party library and roll your own device picker UI. meshcore.js (JavaScript) meshcore-open (Dart) meshcore_py (Python) They have to use magic numbers like rather than referring to constants defined in some authoritative location. None of them have automated tests for their parsers. They’re dragging unnecessary low-level work into high-level languages. For example, everyone is storing and variables. That’s an artifact of the C implementation, where arrays don’t know their size. You don’t have to manually track an array’s size in languages like JavaScript, Python, or Dart. They don’t check data carefully, so they’ll happily pass on garbage data like a negative path length or GPS coordinates that are outside of Earth’s bounds. They all ignore the flags field even though the flags are supposed to indicate which fields are populated . Or at least they’re supposed to in the peer-to-peer messages. For device-to-client messages, they seem to be meaningless. https://codeberg.org/mtlynch/libmeshcore-client Zig does not yet compile to xtensa architecture, which most of the MeshCore devices use. PlatformIO, which most of the MeshCore firmware projects use, does not support Zig. Dart’s ffigen would maybe work with Zig since Zig supports C’s ABI, but it was hard even getting it to work with C. Ditto for Python’s cffi . I got most of the way through writing two new chapters of Refactoring English . I got most of the way through writing the design doc for my photo sharing app idea. I published “The Most Popular Blogs of Hacker News in 2025.” I created my first Flutter app . I created my first cross-language library . I made some contributions to MeshCore meshcore.js . Most of which, the maintainers are ignoring. Minimize in-flight projects AI makes it easier than ever to start new projects, but I’m still the bottleneck on turning them into something production-ready. The result is that I have a lot of projects that are in-flight and waiting for me to review them before I publish them. There’s mental overhead in so much context-switching and task tracking. Publish three chapters of Refactoring English . Publish my 2025 annual review (year 8).

Dart Writing Career Business C

▲ 0

0 views

Stone Tools 6 months ago

XPER on the Commodore 64

In 1984, Gary Kildall and Stewart Chiefet covered "The Fifth Generation" of computing and spoke with Edward Feigenbaum, the father of "expert systems." Kildall started the show saying AI/expert systems/knowledge-based systems (it's all referred to interchangeably) represented a "quantum leap" in computing. "It's one of the most promising new softwares we see coming over the horizon." One year later, Kildall seemed pretty much over the whole AI scene. In an episode on "Artificial Intelligence" he did nothing to hide his fatigue from the guests, "AI is one of those things that people are pinning to their products now to make them fancier and to make them sell better." He pushed back hard against the claims of the guests, and seemed less-than-impressed with an expert system demonstration. The software makers of those "expert systems" begged to differ. There is a fundamental programmatic difference in the implementation of expert systems which enables a radical reinterpretation of existing data, they argued. Guest Dr. Hubert Dreyfus re-begged to re-differ, suggesting it should really be called a "competent system." Rules-based approaches can only get you about 85% of the way toward expertise; it is intuition which separates man from machine, he posited. I doubt Dreyfus would have placed as high as 85% competence on a Commodore 64. The creator of XPER , Dr. Jacques Lebbe, was undeterred, putting what he knew of mushrooms into it to democratize his knowledge. XPER , he reasoned, could do the same for other schools of knowledge even on humble hardware. So, just how much expertise can one cram into 64K anyway? So, what is an "expert system" precisely? According to Edward Feigenbaum, creator of the first expert system DENDRAL, in his book The Fifth Generation: Artificial Intelligence and Japan's Computer Challenge to the World , "It is a computer program that has built into it the knowledge and capability that will allow it to operate at the expert's level. (It is) a high-level intellectual support for the human expert." That's a little vauge, and verges on over-promising. Let's read on. "Expert systems operate particularly well where the thinking is mostly reasoning, not calculating - and that means most of the world's work." Now he's definitely over-promising. After going through the examples of expert systems in use, it boils down to a system which can handle combinatorial decision trees efficiently. Let's look at an example. A doctor is evaluating a patient's symptoms. A way to visualize her thought process for a diagnosis might take the below form. An expert system says, "That looks like a simple decision tree. I happen to know someone who specializes in things like that, hint hint." XPER is a general-purpose tool for building such a tree from expert knowledge, carrying the subtle implication (hope? prayer?) that some ephemeral quality of the decision making process might also be captured as a result. Once the tree is captured, it is untethered from the human expert and can be used by anyone. XPER claims you can use it to build lots of interesting things. It was created to catalog mushrooms, but maybe you want to build a toy. How about a study tool for your children? Let's go for broke and predict the stock market! All are possible , though I'm going to get ahead of your question and say one of those is improbable . I have a couple of specific goals this time. First, the tutorial is a must-do, just look at this help menu. This is the program trying to HELP ME . After I get my head around that alphabet soup, I want to build a weather predictor. The manual explicitly states it as a use case and by gum I'ma gonna do it. I'm hoping that facts like, "Corns ache something intolerable today" and "Hip making that popping sound again" can factor into the prediction at some point. First things first, what does this program do? I don't mean in the high-level, advertising slogan sense, I mean "What specific data am I creating and manipulating with XPER ?" It claims "knowledge" but obviously human knowledge will need to be molded into XPER knowledge somehow. Presently, we don't speak the same language. XPER asks us to break our knowledge down into three discrete categories, with the following relationships of object, feature, and attribute: My Gen-X BS alarm is ringing that something's not fully formed with this method for defining knowledge. Can everything I know really be broken down into three meager components and simple evaluations of their qualities? Defining objects happens in a different order than querying, which makes it a little fuzzy to understand how the two relate. While we define objects as collections of attributes, when querying against attributes to uncover matching objects. The program is well-suited to taxonomic identification. Objects like mushrooms and felines have well-defined, observable attributes that can be cleanly listed. A user of the system could later go through attribute lists to evaluate, "If a feline is over 100kg , has stripes , and climbs trees which feline might it be?" For a weather predictor, I find it difficult to determine what objects I should define. My initial thought was to model "a rainy day" but that isn't predictive. What I really want is to be able to identify characteristics which lead into rainy days. "Tomorrow's rain" is an attribute on today's weather, I have naively decided. This is getting at the heart of what XPER is all about; it is a vessel to hold data points. Choosing those data points is the real work, and XPER has nothing to offer the process. This is where the manual really lets us down. In the Superbase 64 article, I noted how the manual fails by not explaining how to transition from the "old way" to the "new way" of data cataloging. For a program which suggests building toys from it, the XPER manual doesn't provide even a modicum of help in understanding how to translate my goals into XPER objects. The on-disk tutorial database of "felines" shows how neatly concepts like "cat identification" fit into XPER framing. Objects are specific felines like "jaguar," "panther," "mountain lion." Features suggest measureable qualities like "weight," "tree climbing," "fur appearance," "habitat" etc. Attributes get more specific, as "over 75kg," "yes," "striped," and "jungle." For the weather predictor, the categories of data are similarly precise. "cloud coverage," "temperature," "barometer reading," "precipitation," "time of year," "location," and so forth may serve our model. Notice that for felines we could only define rough ranges like "over 75kg" and not an exact value. We cannot set a specific weight and ask for "all cats over some value." XPER contains no tools for "fuzzy" evaluations and there is no way to input continuous data. Let's look at the barometer reading, as an example. Barometer data is hourly, meaning 24 values per day. How do I convert that into a fixed value for XPER ? To accurately enter 24 hours of data, I would need to set up hourly barometer features and assign 30? 50? possible attributes for the barometer reading. Should we do the same for temperature? Another 24 features, each with 30 or more attributes, one per degree change? Precipitation? Cloud coverage? Wind speed and direction? Adorableness of feline? Besides the fact that creating a list of every possible barometric reading would be a ridiculous waste of time, it's not even possible in the software. A project is limited to We must think deeply about what data is important to our problem, and I'd say that not even the expert whose knowledge is being captured would know precisely how to structure XPER for maximum accuracy. The Fifth Generation warns us: "GiT GuD" as the kids say. ( Do they still say that?! ) The graphic above, output by XPER 's "Printer" module, reveals the underlying data structure of the program. Its model of the data is called a "frame," a flat 2-D graph where objects and attributes collide. That's it. Kind of anticlimactic I suppose, but it imbues our data with tricks our friend Superbase can't perform. First, this lets us query the data in human-relatable terms, as a kind of Q&A session with an expert. "Is it a mammal?" "Does it have striped fur?" "Does it go crazy when a laser pointer shines at a spot on the floor?" Through a session, the user is guided toward an object, by process of elimination, which matches all known criteria, if one exists. Second, we can set up the database to exclude certain questions depending upon previous answers. "What kind of fur does it have?" is irrelevant if we told it the animal is a fish, and features can be set up to have such dependencies. This is called a father/son relationship in the program, and also a parent/child relationship in the manual. "fur depends on being a mammal," put simply. Third, we can do reverse queries to extract new understandings which aren't immediately evident. In the feline example it isn't self-evident, but can be extracted, that "all African felines which climb trees have retractile claws." For the weather predictor I hope to see if "days preceding a rainy day" share common attributes. The biggest frustration with the system is how all knowledge is boxed into the frame. For the weather predictor, this is frustrating. With zero relationship between data points, trends cannot be identified. Questions which examine change over time are not possible, just "Does an object have an attribute, yes or no?" To simulate continuous data, I need to pre-bake trends of interest into each object's attributes. For example, I know the average barometric pressure for a given day, but because XPER can't evaluate prior data, it can't evalute if the pressure is rising or falling. Since it can't determine this for itself, I must encode that as a feature like "Barometric Trend" with attributes "Rising," "No Change," and "Falling." The more I think about the coarseness with which I am forced to represent my data, the more clear it is to me how much is being lost with each decision. That 85/15 competency is looking more like 15/85 the other direction. Collecting data for the weather predictor isn't too difficult. I'm using https://open-meteo.com to pull a spreadsheet on one month of data. I'll coalesce hourly readings, like barometric pressure, into average daily values. Temperature will be a simple "min" and "max" for the day. Precipitation will be represented as the sum of all precipitation for the day. And so on. As a professional not-a-weather-forecaster, I'm pulling whatever data strikes me as "interesting." In the spirit of transparency, I mentally abandoned the "expert" part of "expert system" pretty early on. This guy *points at self with thumbs* ain't no expert. Having somewhat hippy-dippy parents, I've decided that Mother Earth holds secrets which elude casual human observation. To that end, I'm including "soil temperature (0 - 7cm)" as a data point, along with cloud coverage, and relative humidity to round out my data for both systematic and "I can't spend months of my life on this project" reasons. After collecting November data for checkpoint years 2020, 2022, and 2024, actually entering the data is easier than expected. XPER provides useful F-Key shortcuts which let me step through objects and features swiftly. What I thought would take days to input wound up only being a full afternoon. Deciding which data I want, collecting it, and preparing it for input was the actual work, which makes sense. Entering data is easy; becoming the expert is hard. Even as I enter the data, I catch fleeting glimpses of patterns emerging and they're not good. It's an interesting phenomenon, having utterly foreign data start to feel familiar. Occasionally I accidentally correctly predict if the next day's weather has rain. Am I picking up on some subliminal pattern? If so, might XPER "see" what I'm seeing? I'm not getting my hopes up, but I wonder if early fascination with these forays into AI was driven by a similar feeling of possibility? We're putting information into a system and legimitately not knowing what will come out of the processing. There is a strong sense of anticipation; a powerful gravity to this work. It is easy to fool myself into believing I'm unlocking a cheat code to the universe. Compare this to modern day events if you feel so inclined. At the same time, there's obviously not enough substance to this restricted data subset. As I enter that soil temperature data, 90% of the values keep falling into the same bin. My brainstorm for this was too clever by half, and wrong. As well, as I enter data I find sometimes that I'm entering exactly the same information twice in a row, but the weather results are different enough as to make me pause. Expert systems have a concept of "discriminating" and "non-discriminating" features. If a given data point for every object in a group of non-eliminated objects is the same, that data point is said to be "non-discriminating." In other words, "it don't matter" and will be skipped by XPER during further queries. The question then is, whose fault is this? Did I define my data attributes incorrectly for this data point or is the data itself dumb? I can only shrug, "Hey, I just work here." XPER has a bit of a split personality. Consider how a new dataset is created. From the main menu, enter the Editor. From there you have four options. First, I go to option for the seemingly redundantly named "Initializing Creating." Then I set up any features, attributes, and objects I know about, return to this screen, and save with option . Later I want to create new objects. I type for "Creating" and am asked, "Are you sure y/n" Am I sure ? Am I sure about what ? I don't follow, but yes, I am sure I want to create some objects. I hit and I'm back at a blank screen, my data wiped. That word "initializing" is doing the heavy lifting on this menu. "Initialize" means "first time setup of a dataset," which also allows, almost as a side effect, the user an opportunity to input whatever data happens to be known at that moment . "Initial Creation" might be more accurate? Later, when you want to add more data, that means you now want to edit your data, called "revising" in XPER , and that means option . Option is only ever used the very first time you start a new data set. is for every time you append/delete afterward. The prompts and UI are unfortunately obtuse and unhelpful. "Are you sure y/n" is too vague to make an informed decision. The program would benefit greatly from a status bar displaying the name of the current in-memory dataset, if it has been saved or not, and a hint on how close we are to the database limit. Prompts should be far more verbose, explaining intent and consequence. A status bar showing the current data set would be especially useful because of the other weird quirk of the program: how often it dumps data to load in a new part of the program. XPER is four independent programs bound together by a central menu. Entering a new area of the program means effectively loading a new program entirely, which requires its own separate data load. If you see the prompt "Are you sure y/n" what it really means is, "Are you sure you want me to forget your data because the next screen you go to will not preserve it. y/n" That's wordy, but honest. With that lesson learned, I'm adding three more data points to the weather predictor: temperature trend, barometric trend, and vapor pressure deficit (another "gut feeling" choice on my part). Trends should make up a little for the lack of continuous data. This will give me a small thread of data which leads into a given day, the data for that day, and a little data leading out into the next day. That fuzzes up the boundaries. It feels right, at the very least. Appending the new information is easy and painless. Before, I used F3/F4 to step through all features of a given object. This time I'm using F5/F6 to step through a single feature across all objects. This only took about fifteen minutes. I'm firmly in "manual memory management" territory with this generation of hardware. Let's see where we sit relative to the maximum potential. Features like this really makes one appreciate the simple things in life like a mouse, gui checklists, and simple grouping mechanisms. XPER can compare objects or groups of objects against one another, identifying elements which are unique to one group or the other. You get two groups, full stop. Items in those groups and only those groups will be compared when using the command. We can put objects individually into one of those two groups, or we can create an object definition and request that "all objects matching this definition go into group 1 or 2". This is called a STAR object. I created two star objects: one with tomorrow's weather as rain, and one with tomorrow's weather as every type except rain. Groups were insta-built with the simple command where means and means , my "rainy day" star object. I can ask for an AND or OR comparison between the two groups, and with any luck some attribute will be highlighted (invert text) or marked (with ) as being unique or exclusive to one group or the other. If we find something, we've unlocked the secret to rain prediction! Take THAT, Cobra Commander ! Contrary to decades of well-practiced Gen-X cynicism, I do feel a tiny flutter of "hope" in my stomach. Let's see what the XPER analysis reveals! The only thing unique between rainy days and not is the fact that it rained. The Jaccard Distance , developed by Grove Karl Gilbert in 1884, is a measure of the similiarity/diversity between two sets (as in "set theory" sets). The shorter the distance, the more "similar" the compared sets are. XPER can measure this distance between objects. where is the object ID of interest, will run a distance check of that object against all other objects. On my weather set with about 90 objects, it took one minute to compare Nov. 1, 2020 with all other days at 100% C64 speed. Not bad! What can we squeeze out of this thing? By switching into "Inquirer" mode, then loading up the data set of interest, a list of top level object features are presented. Any features not masked by a parent feature are "in the running" as possible filters to narrow down our data. So, we start by entering what we know about our target subject. One by one, we fill in information by selecting a feature then selecting the attribute(s) of that feature, and the database updates its internal state, quietly eliminating objects which fall outside our inquiry. The command will look at the "remaining objects," meaning "objects which have not yet been eliminated by our inquiry so far." With the command as in to run it against the "jaguar" we can ask XPER to tell us which features, in order, should we answer to narrrow down to the jaguar as quickly as possible. It's kind of ranking the features in order of importance to that specific object. It sounds a bit like feature weighting , but it's definitely not. XPER isn't anywhere close to that level of sophistication. In this data set, if I answer "big" for "prey size" I immediately zero in on the jaguar, it being the currently most-discriminating feature for that feline. You might be looking at this and wondering how, exactly, this could possibly predict the weather. You and me, both, buddy. The promise of Fifth Gen systems and the reality are colliding pretty hard now. Feigenbaum and "The Fifth Generation" have been mentioned a few times so far, so I should explain that a little. Announced in 1981, started in 1982, and lasting a little more than a decade, "The Fifth Generation" of computing was Japan's moniker for an ambitious nationwide initiative. According to the report of Japan's announcement, Fifth Generation Computer Systems : Proceedings of the International Conference on Fifth Generation Computer Systems, Tokyo, Japan, October 19-22, 1981, Japan had four goals: In Fifth Generation Computers: Concepts, Implementation, and Uses (1986), Peter Bishop wrote, "The impact on those attending the conference was similar to that of the launch of the Soviet Sputnik in 1957." During a hearing before the Committee on Science and Technology in 1981, Representative Margaret Heckler said , "When the Soviets launched Sputnik I, a remarkable engineering accomplishment, the United States rose to the challenge with new dedication to science and technology. Today, our technology lead is again being challenged, not just by the Soviet Union, but by Japan, West Germany, and others." Scott Armstrong writing for The Christian Science Monitor in 1983, in an article titled, "Fuchi - Japan's computer guru" said, "The debate now - one bound to intensify in the future - is whether the US needs a post-Sputnik-like effort to counter the Japanese challenge. Japan's motive (reflects) a sense of nationalism as much as any economic drive." Innovation Policies for the 21st Century: Report of a Symposium (2007) remarked of Japan's Fifth Generation inroads into supercomputers, "This occasioned some alarm in the United States, particularly within the military." It would be fair to say there was "Concern," with a capital C. In 1989's The Fifth Generation: The Future of Computer Technology by Jeffrey Hsu and Joseph Kusnan (separate from Feigenbaum's The Fifth Generation ) said Japan had three research projects The "Fifth Generation" was specifically the software side which the conference claimed, "will be knowledge information processing systems based on innovative theories and technologies that can offer the advanced functions expected to be required in the 1990's overcoming the technical limitations inherent in conventional computers." Expert systems played a huge role during the AI boom of the 80s, possibly by distancing itself from "AI" as a concept, focusing instead on far more plausible goals. It's adjacent to, but isn't really, "artificial intelligence." This Google N-Gram chart shows how "expert system" had more traction than the ill-defined "artificial intelligence." Though they do contain interesting heuristics, there is no "intelligence" in an expert system. Even the state of the art demonstrated on Computer Chronicles looked no more "intelligent" than a Twine game . That sounds non-threatening; I don't think anyone ever lost a job to a Choose Your Own Adventure book. In those days, even something that basic had cultural punch. Feigenbaum's The Fifth Generation foreshadowed today's AI climate, if perhaps a bit blithely. That guy wasn't alone. In 1985, Aldo Cimino, of Campbell's Soup Co., had his 43 years of experience trouble-shooting canned soup sterilizers dumped onto floppy by knowledge engineers before he retired. They called it "Aldo on a Disk" for a time. He didn't mind, and made no extra money off the brain dump, but said the computer "only knows 85% of what he does." Hey, that's the same percentage Hubert Dreyfus posited at the start of this article! That system was retired about 10 years later, suffering from the same thing that a lot of expert systems of the day did: brittleness. From the paper, "Expert Systems and Knowledge-Based Engineering (1984-1991)" by Jo Ann Oravec, "Brittleness (inability of the system to adapt to changing conditions and input, thus producing nonsensical results) and “knowledge engineering bottlenecks” were two of the more popular explanations why early expert system strategies have failed in application." Basically, such systems were inflexible to changing inputs (that's life), and nobody wanted to spend the time or money to teach them the new rules. The Campbell's story was held up as an exemplar of the success possible with such systems, and even it couldn't keep its job. It was canned. (Folks, jokes like that are the Stone Tools Guarantee™ an AI didn't write this.) Funnily enough, the battles lost during the project may have actually won the war. There was a huge push toward parallelism in compute during this period. You might be familiar with a particularly gorgeous chunk of hardware called the Connection Machine. Japan's own highly parallel computers, the Parallel Inference Machines (PIM), running software built with their own bespoke programming language, KL1, seemed like the future. Until it didn't. PIM and Thinking Machines and others all fell to the same culprit. Any gains enjoyed by parallel systems were relatively slight and the software to take advantage of those parallel processors was difficult to write. In the end the rise of fast, cheap CPUs evaporated whatever advantages parallel systems promised. Today we've reversed course once more on our approach to scaling compute. As Wikipedia says, "the hardware limitations foreseen in the 1980s were finally reached in the 2000s" and parallelism became fashionable once more. Multi-core CPUs and GPUs with massive parallelism are now put to use in modern AI systems, bringing Fifth Generation dreams closer to reality 35 years after Japan gave up. In a "Webster's defines an expert system as..." sense, I suppose XPER meets a narrow definition. It can store symbolic knowledge in a structured format and allow non-experts to interrogate expert knowledge and discover patterns within. That's not bad for a Commodore 64! If we squint, it could be mistaken for a "real" expert system at a distance, but it's not a "focuser ." It borrows the melody of expert systems, yet is nowhere near the orchestral maneuverings of its true "fifth generation" brothers and sisters. Because XPER lacks inference, the fuzzy result of inquiry relies on the human operator to make sense of it. Except for mushroom and feline taxonomy, you're unlikely to get a "definitive answer" to queries. Rather, the approach is to put in data and hope to narrow the possibility space down enough to have something approachable. Then, look through that subset and see if a tendency can be inferred. The expert was in our own hearts all along. Before I reveal the weather prediction results, we must heed an omen from page 10 of the manual. I'm man enough to admit my limits: I'm a dummy. When a dummy feeds information into XPER , the only possible result is that XPER itself also becomes a dummy. With that out of the way, here's a Commodore 64, using 1985's state-of-the-art AI expert system, predicting tomorrow's weather over two weeks. Honestly, not a lot. Ultimately, this wound up being far more "toy" than "productivity," much to my disappointment. A lot of that can be placed on me, for not having an adequate sense of the program's limitations going in. Some of that's on XPER though, making promises it clearly can't keep. Perhaps that was pie-in-the-sky thinking, and a general "AI is going to change everything" attitude. Everyone was excited for the sea-change! It was used for real scientific data analysis by real scientists, so it would be very unfair of me to dismiss it entirely. On the other hand, there were contemporary expert systems on desktop microcomputers which provided far more robust expert system implementations and the advantages of those heuristic evaluations. In that light, XPER can't keep up though it is a noble effort. Overall, I had fun working with it. I honestly enjoyed finding and studying the data, and imagining what could be accomplished by inspecting it just right. Notice that XPER was conspicuously absent during that part of that process, though. Perhaps the biggest takeaway is "learning is fun," but I didn't need XPER to teach me that. Ways to improve the experience, notable deficiencies, workarounds, and notes about incorporating the software into modern workflows (if possible). VICE, in C64 mode Speed: ~200% (quirk noted later) Snapshots are in use; very handy for database work Drives: XPER seems to only support a single drive XPER v1.0.3 (claims C128, but that seems to be only in "C64 Mode") 250 objects 50 features 300 attributes (but no more than 14 in any given feature) To increase productivity in low-productivity areas. To meet international competition and contribute toward international cooperation. To assist in saving energy and resources. To cope with an aged society. Superspeed Computer Project (the name says it all) The Next-Generation Industries Project (developing the industrial infrastructure to produce components for a superspeed computer) Fifth-Generation Computer Project Any time we're in C64 land, disk loading needs Warp mode turned on. Actual processing of data is fairly snappy, even at normal 100% CPU speed; certainly much faster than Superbase . I suspect XPER is mostly doing bitwise manipulations, nothing processing intensive. XPER did crash once while doing inquiry on the sample feline database. Warp mode sometimes presents repeating input, sometimes not. I'm not entirely certain why it was inconsistent in that way. XPER seems to only recognize a single disk drive. Don't even think about it, you're firmly in XPER Land. XPER 2 might be able to import your data, though. You'll still be in XPER Land, but you won't be constrained by the C64 any longer. As a toy, it's fine. For anything serious, it can't keep up even with its contemporaries: No fuzzy values No weighted probabilities/certainties No forward/backward chaining Limited "explanation" system, as in "Why did you choose that, XPER ?" (demonstrated by another product in Computer Chronicles 1985 "Artificial Intelligence" episode) No temporal sequences (i.e. data changes over time) No ability to "learn" or self-adapt over time No inference

Dart Hardware AI

▲ 0

0 views

Max Bernstein 6 months ago

The GDB JIT interface

GDB is great for stepping through machine code to figure out what is going on. It uses debug information under the hood to present you with a tidy backtrace and also determine how much machine code to print when you type . This debug information comes from your compiler. Clang, GCC, rustc, etc all produce debug data in a format called DWARF and then embed that debug information inside the binary (ELF, Mach-O, …) when you do or equivalent. Unfortunately, this means that by default, GDB has no idea what is going on if you break in a JIT-compiled function. You can step instruction-by-instruction and whatnot, but that’s about it. This is because the current instruction pointer is nowhere to be found in any of the existing debug info tables from the host runtime code, so your terminal is filled with . See this example from the V8 docs: Fortunately, there is a JIT interface to GDB. If you implement a couple of functions in your JIT and run them every time you finish compiling a function, you can get the debugging niceties for your JIT code too. See again a V8 example: Unfortunately, the GDB docs are somewhat sparse . So I went spelunking through a bunch of different projects to try and understand what is going on. GDB expects your runtime to expose a function called and a global variable called . GDB automatically adds its own internal breakpoints at this function, if it exists. Then, when you compile code, you call this function from your runtime. In slightly more detail: This is why you see compiler projects such as V8 including large swaths of code just to make object files: Because this is a huge hassle, GDB also has a newer interface that does not require making an ELF/Mach-O/…+DWARF object. This new interface requires writing a binary format of your choice. You make the writer and you make the reader. Then, when you are in GDB, you load your reader as a shared object. The reader must implement the interface specified by GDB : The function pointer does the bulk of the work and is responsible for matching code ranges to function names, line numbers, and more. Here are some details from Sanjoy Das . Only a few runtimes implement this interface. Most of them stub out the and function pointers: I think it also requires at least the reader to proclaim it is GPL via the macro . Since I wrote about the perf map interface recently, I have it on my mind. Why can’t we reuse it in GDB? I suppose it would be possible to try and upstream a patch to GDB to support the Linux perf map interface for JITs. After all, why shouldn’t it be able to automatically pick up symbols from ? That would be great baseline debug info for “free”. In the meantime, maybe it is reasonable to create a re-usable custom debug reader: It would be less flexible than both the DWARF and custom readers support: it would only be able to handle filename and code region. No embedding source code for GDB to display in your debugger. But maybe that is okay for a partial solution? Update: Here is my small attempt at such a plugin. V8 notes in their GDB JIT docs that because the JIT interface is a linked list and we only keep a pointer to the head, we get O(n 2 ) behavior. Bummer. This becomes especially noticeable since they register additional code objects not just for functions, but also trampolines, cache stubs, etc. Since GDB expects the code pointer in your symbol object file not to move, you have to make sure to have a stable symbol file pointer and stable executable code pointer. To make this happen, V8 disables its moving GC. Additionally, if your compiled function gets collected, you have to make sure to unregister the function. Instead of doing this eagerly, ART treats the GDB JIT linked list as a weakref and periodically removes dead code entries from it. Compile a function in your JIT compiler. This gives you a function name, maybe other metadata, an executable code address, and a code size Generate an entire ELF/Mach-O/… object in-memory (!) for that one function, describing its name, code region, maybe other DWARF metadata such as line number maps Write a linked list node that points at your object (“symfile”) Link it into the linked list Call , which gives GDB control of the process so it can pick up the new function’s metadata Optionally, break into (or crash inside) one of your JITed functions At some point, later, when your function gets GCed, unregister your code by editing the linked list and calling again CoreCLR/.NET JavaScriptCore ART which looks like it does something smart about grouping the JIT code entries together ( ), but I’m not sure exactly what it does TomatoDotNet a minimal example It looks like Dart used to have support for this but has since removed it yk write yk read asmjit-utilities write asmjit-utilities read Erlang/OTP write Erlang/OTP read FEX write FEX read buxn-jit write buxn-jit read box64 write box64 read When registering code, write the address and name to as you normally would Write the filename as the symfile (does this make the magic number?) Have the debug info reader just parse the perf map file

Dart Open Source Erlang Programming

▲ 0

0 views

matklad 6 months ago

The Second Great Error Model Convergence

I feel like this has been said before, more than once, but I want to take a moment to note that most modern languages converged to the error management approach described in Joe Duffy’s The Error Model , which is a generational shift from the previous consensus on exception handling. C++, JavaScript, Python, Java, C# all have roughly equivalent , , constructs with roughly similar runtime semantics and typing rules. Even functional languages like Haskell, OCaml, and Scala feature exceptions prominently in their grammar, even if their usage is frowned upon by parts of the community. But the same can be said about Go, Rust, Swift, and Zig! Their error handling is similar to each other, and quite distinct from the previous bunch, with Kotlin and Dart being notable, ahem, exceptions. Here are some commonalities of modern error handling: First , and most notably, functions that can fail are annotated at the call side. While the old way looked like this: the new way is There’s a syntactic marker alerting the reader that a particular operation is fallible, though the verbosity of the marker varies. For the writer, the marker ensures that changing the function contract from infallible to fallible (or vice versa) requires changing not only the function definition itself, but the entire call chain. On the other hand, adding a new error condition to a set of possible errors of a fallible function generally doesn’t require reconsidering rethrowing call-sites. Second , there’s a separate, distinct mechanism that is invoked in case of a detectable bug. In Java, index out of bounds or null pointer dereference (examples of programming errors) use the same language machinery as operational errors. Rust, Go, Swift, and Zig use a separate panic path. In Go and Rust, panics unwind the stack, and they are recoverable via a library function. In Swift and Zig, panic aborts the entire process. Operational error of a lower layer can be classified as a programming error by the layer above, so there’s generally a mechanism to escalate an erroneous result value to a panic. But the opposite is more important: a function which does only “ordinary” computations can be buggy, and can fail, but such failures are considered catastrophic and are invisible in the type system, and sufficiently transparent at runtime. Third , results of fallible computation are first-class values, as in Rust’s . There’s generally little type system machinery dedicated exclusively to errors and expressions are just a little more than syntax sugar for that little Go spell. This isn’t true for Swift, which does treat errors specially. For example, the generic function has to explicitly care about errors, and hard-codes the decision to bail early: Swift does provide first-classifier type for errors. Should you want to handle an exception, rather than propagate it, the handling is localized to a single throwing expression to deal with a single specific errors, rather than with any error from a block of statements: Swift again sticks to more traditional try catch, but, interestingly, Kotlin does have expressions. The largest remaining variance is in what the error value looks like. This still feels like a research area. This is a hard problem due to a fundamental tension: The two extremes are well understood. For exhaustiveness, nothing beats sum types ( s in Rust). This I think is one of the key pieces which explains why the pendulum seemingly swung back on checked exceptions. In Java, a method can throw one of the several exceptions: Critically, you can’t abstract over this pair. The call chain has to either repeat the two cases, or type-erase them into a superclass, losing information. The former has a nasty side-effect that the entire chain needs updating if a third variant is added. Java-style checked exceptions are sensitive to “N to N + 1” transitions. Modern value-oriented error management is only sensitive to “0 to 1” transition. Still, if I am back to writing Java at any point, I’d be very tempted to standardize on coarse-grained signature for all throwing methods. This is exactly the second well understood extreme: there’s a type-erased universal error type, and the “throwableness” of a function contains one bit of information. We only care if the function can throw, and the error itself can be whatever. You still can downcast dynamic error value handle specific conditions, but the downcasting is not checked by the compiler. That is, downcasting is “save” and nothing will panic in the error handling mechanism itself, but you’ll never be sure if the errors you are handling can actually arise, and whether some errors should be handled, but aren’t. Go and Swift provide first-class universal errors, like Midori. Starting with Swift 4, you can also narrow the type down. Rust doesn’t really have super strong conventions about the errors, but it started with mostly enums, and then and shone spotlight on the universal error type. But overall, it feels like “midpoint” error handling is poorly served by either extreme. In larger applications, you sorta care about error kinds, and there are usually a few place where it is pretty important to be exhaustive in your handling, but threading necessary types to those few places infects the rest of the codebases, and ultimately leads to “a bag of everything” error types with many “dead” variants. Zig makes an interesting choice of assuming mostly closed-world compilation model, and relying on cross-function inference to learn who can throw what. What I find the most fascinating about the story is the generational aspect. There really was a strong consensus about exceptions, and then an agreement that checked exceptions are a failure , and now, suddenly, we are back to “checked exceptions” with a twist, in the form of “errors are values” philosophy. What happened between the lull of the naughts and the past decade industrial PLT renaissance? On the one hand, at lower-levels you want to exhaustively enumerate errors to make sure that: internal error handling logic is complete and doesn’t miss a case, public API doesn’t leak any extra surprise error conditions. On the other hand, at higher-levels, you want to string together widely different functionality from many separate subsystems without worrying about specific errors, other than: separating fallible functions from infallible, ensuring that there is some top-level handler to show a 500 error or an equivalent.

C++ Programming Backend C# Dart

▲ 0

0 views

Oya Studio 7 months ago

Better than JSON

An in-depth look at why Protobuf can outperform JSON for modern APIs, with practical Dart examples showing how strong typing, binary serialization, and shared schemas improve both performance and developer experience.

JSON Dart Performance API

▲ 0

0 views

Oya Studio 7 months ago

Vibe Coding: Beyond the Joke – A Serious Tool for Rapid Prototyping

Learn how to create custom animated scenes for your live streams using Flutter as a source in OBS.

Tutorial Dart

▲ 0

0 views

Ivan Sagalaev 7 months ago

Pet project restart

So what happened was, I have developed my shopping list to the point where it got useful to me , after which I lost interest in working on it. You know, the usual story… It was however causing me enough annoyances to still want to get back to it eventually. So a few weeks ago, after not having done any programming for a year, I finally broke through the dread of launching my IDE again and started on slowly fixing the accumulated bitrot. And through the last several days I was on a blast implementing some really useful stuff and feeling the familiar thrill of being in the flow . Since I was mostly focused on making the app useful I didn't pay a lot of attention to the UI, so most of the annoyances were caused purely by my not wanting to spend much time on fighting Android APIs. Here's one of those. The app keeps several shopping lists in a swipe-able pager, and at the same time swiping is how you remove items from the list while going through the store. The problem was that swiping individual items was really sensitive to a precise finger movement, so instead it would often be intercepted by the pager and it would switch to the next list instead. That's fixed now (with an ugly hack). But the biggest deficiency of the app was that it didn't let me get away from one particular grocery store that I started to rather dislike. You might find it weird that some app could exert such control over my actions, but let me explain. It all comes down to three missing features… The central feature of my app is remembering the order in which I buy grocery items. This means I need a separate list for every store, as every one of them has a different physical layout. By the time I was thinking of switching to another store I already had an idea about a new evolution of the order training algorithm in the app, and a new store would be a great dogfooding use case for it. So I've got a sort of mental block: I didn't want to switch stores before I implemented this new algorithm. Over some years of using the app with a single store I've been manually associating grocery categories with products ("dairy", "produce", etc.). They are color coded, which make the list easier to scan visually. But starting a new list for another store meant that I would either need to do it all again for every single item, or accept looking at a dull, unhelpful gray list. What I really needed was some smart automatic prediction, but I didn't have it. I usually collect items in a list over a week for an upcoming visit to the store, and sometimes I realize that I need something that it simply doesn't carry, or my other errands would make it easier to go to another store. At this point I'd like to select all the items in a filled-up list and move them to another, which the app also couldn't do. See, it all makes sense! Now, of course it wasn't a literal impossibility for me to go to other stores, and on occasion I did, it just wasn't very convenient. But these are all pretty major deficiencies, and I'm not ready to offer the app to other people without them sorted out. Anyway… Over the course of three weeks I implemented two of those big features: category guessing and cross-store moves. And I convinced myself that I can live with the old ordering algorithm for a while. So now I can finally wean myself off of the QFC on Redmond Way (which keeps getting worse, by the way) and start going to another QFC (a completely different experience). All the categories (item colors) you see in the screencaps above were guessed automatically. My prediction model works pretty well on my catalog of 400+ grocery items: the data comes from me tagging them manually while doing my own shopping these past 4 years. And this also means, of course, that it's biased towards what I tend to buy. It doesn't know much about alcohol or frozen ready-to-eat foods, for example. I'm planning to put up a little web app to let other people help me train it further. I'll keep y'all posted! One important note though… No, it's not a frigging LLM! It's technically not even ML , as there is no automatic calibration of weights in a matrix or anything. Instead it's built on a funny little trick I learned at Shutterstock while working on a search suggest widget. I'll tell you more when I launch the web app. When I started developing the app, I used the official UI toolkit documented on developer.android.com. It's a bunch of APIs with a feel of a traditional desktop GUI paradigm (made insanely complicated by Google "gurus"). Then the reactive UI revolution happened, and if you wanted something native for Android, it was represented by Flutter . Now they're recommending Compose . I'm sure both are much better than the legacy APIs, but I'm kind of happy I wasn't looking in this space for a few years and wasn't tempted to rewrite half the code. Working in the industry made me very averse to constant framework churn. I'm not making any promises, but as the app is taking shape rather nicely, I'm again entertaining the idea of actually… uhm… finishing it. Which would mean beta testing, commissioning professional artwork and finally selling the final product. The central feature of my app is remembering the order in which I buy grocery items. This means I need a separate list for every store, as every one of them has a different physical layout. By the time I was thinking of switching to another store I already had an idea about a new evolution of the order training algorithm in the app, and a new store would be a great dogfooding use case for it. So I've got a sort of mental block: I didn't want to switch stores before I implemented this new algorithm. Over some years of using the app with a single store I've been manually associating grocery categories with products ("dairy", "produce", etc.). They are color coded, which make the list easier to scan visually. But starting a new list for another store meant that I would either need to do it all again for every single item, or accept looking at a dull, unhelpful gray list. What I really needed was some smart automatic prediction, but I didn't have it. I usually collect items in a list over a week for an upcoming visit to the store, and sometimes I realize that I need something that it simply doesn't carry, or my other errands would make it easier to go to another store. At this point I'd like to select all the items in a filled-up list and move them to another, which the app also couldn't do.

Mobile Android Programming Dart

▲ 0

0 views

Justin Duke 9 months ago

September, 2025

The last of summer's grip finally loosened its hold this September, and Richmond began its annual transformation into something gentler and more contemplative. This morning's walk with Telly required a dusting-off of the closet-buried Patagonia puffer jacket; it's perfect for walks with Lucy, who has graduated into the Big Kid stroller making it easier than ever for her to point at every dog ("dah!"), every bird (also "dah!"), every passing leaf that dared to flutter in her line of sight. As you will read below, the big corporate milestone for me this month was sponsoring Djangocon and having our first offsite over the course of a single week. Sadly, our Seattle trip was once again canceled. Haley and Lucy both got a little sick, and we had to abandon course. It's weird to think this will be the first year since 2011 that we have not stepped foot in the Pacific Northwest. More than anything though, I learned this month for the first time how impossibly difficult it is to be away from your daughter for six days. It is something I hope I have to go through again for a very long time.

Dart Django Career Business

▲ 0

1 views

Anton Zhiyanov 9 months ago

Go is #2 among newer languages

I checked out several programming languages rankings. If you only include newer languages (version 1.0 released after 2010), the top 6 are: ➀ TypeScript, ➁ Go, ➂ Rust, ➃ Kotlin, ➄ Dart, and ➅ Swift. Sources: IEEE , Stack Overflow , Languish . I'm not using TIOBE because their method has major flaws. TypeScript's position is very strong, of course (I guess no one likes JavaScript these days). And it's great to see that more and more developers are choosing Go for the backend. Also, Rust scores very close in all rankings except IEEE, so we'll see what happens in the coming years.

JavaScript Kotlin Programming Dart

▲ 0

0 views

Oya Studio 10 months ago

Flutter safe area is a mess

A friendly deep-dive into why Flutter’s safe area handling often causes more headaches than it solves, and how simpler APIs can sometimes be the better choice.

Mobile Dart

▲ 0

0 views

Oya Studio 10 months ago

Develop apps for Omarchy (with Flutter)

There’s something about discovering a new, beautifully crafted operating system that makes you feel like a kid again. The promise of a lean, polished Linux setup—without the painful hours of configuration—was too tempting. So, there I was in the quiet of the morning, coffee in hand, ready to see what Omarchy was all about and whether I could build apps for it with Flutter.

Dart Tutorial

▲ 0

0 views

annie's blog 11 months ago

Let there be lapses

Let there be lapses Weeds in the garden, unswept porches, A walk never taken, A flower unnoticed, Missed bill, missed text, missed appointment. Let there be undone things Half-written sentences never finished A stack of books never read Blank pages, unseen lines Words never seen or heard or spoken. Let there be glory in what-is-not — All the unachieved Unbelieved Underserved Overlooked. Let us glory in these. Let there be errors Not just the tiny ones we can laugh away But enormous, life-altering errors. Huge risks taken which do not end well. Huge efforts made which result in what we call failure. (In fairness, Any effort is success in certain realities.) But let us — for a moment — judge by the world of machines, Of binaries Of industrialized morality And call it failure. Failure is the word we assign to all unexpected outcomes. So, let there be failure. Let failure warp our seeing and diminish our being, Let it ride among us waving a torch, Shame-blasting and guilt-smearing, Blinding us with ridiculously disproportional fiery judgment, Grinding nose to dirt Binding self to work. Let there be mistakes which make us weep Keep us awake at night Cause us to question our sanity, our decency, Our right to be here, Our ability to keep being here. Let there be broken edges Sawed-off pieces we cannot smooth down Pointy bits irritating and upsetting Dangling splinters and shards over chasms of regret. Let there be surrender. Let us call it what it is: giving up. Surrender sounds too noble, Enlightened, as if I didn’t have to but I chose to. That’s not what this is. Let there be quitting. Let there be Done. Not because we see what we have made, and it is good. This is not putting a bow on a gift. This is saying some things are too broken to be fixed. Let there be giving up. Lay down there, lay down, be still, give up. Face in the mud, breathing in, wheezing in the stuff of life, the dirt, The lowly dirt, the trudged-upon dirt, the worthless dirt From which we came and to which we all return. Let us lay there, breathing in this dirt, This pure self This known self This elemental self. Hell yes, failure. I embrace you. Brother! Sister! Mother! Father! Come quickly! Come and rejoice, for I have failed! Come and celebrate! Set out the feast! Call the guests! And enter into the joy of your child: Humanity raw Humanity broken Humanity dirty Humanity ill-fitted to survive Humanity traumatized Humanity doing such a fucked-up job of it Humanity violent and stumbling Humanity bruised and crusted at the edges Humanity clawing its way from the dark tunnel of history Humanity side-eyeing the stars while blood drips from our fingers Humanity bargaining for the right to squirm Humanity bringing a sword to a gunfight Humanity bullshitting Humanity asking clever little questions Humanity dressed in robes, obsessed with ovaries Humanity unhinged and in charge Humanity waving exasperated hands in the air Humanity dishing out pieces of pie Humanity weeping at the sight of spring flowers Humanity with big rough hands so careful so gentle holding a tiny new fragile thing Humanity with smooth precise hands making deals, ending lives Humanity dropping bombs Humanity being a big dumb bully Humanity the most awkward of the species Humanity voted most likely to secede from the planet Humanity pointing and saying look at this! wow! Humanity wondering, always wondering Humanity exhausted sitting in a patch of sunlight Being dirt. Dirt with form, dirt with spirit. Pale faces float through quiet rooms, ghostly fingers flutter in hallways. Pens move across expensive paper. Golden liquid sloshes in crystal while murmuring voices ooze and wind and hush and tell us there is nothing to worry about.  But this is no time to be civilized.  Let there be lapses: Lapses of courtesy, lapses of decorum. Failures of politeness. Refusals to conform. Let there be a wildness ringing in us for each other — Hissing, bared teeth, spitting — Reverberating, thrumming, cracking the marble palaces full of dead men’s bones.

Writing Java Dart

▲ 0

0 views

Oya Studio 1 years ago

I changed my mind about Vector Graphics

Managing image assets in a Flutter app can quickly become tedious. Learn why vector graphics might not be the silver bullet you thought they were.

Mobile Dart Design

▲ 0

0 views