Pundit ClaimsAI claims by Gary Marcus and Ed Zitron, with the evidence

Ed Zitron Where's Your Ed At Published

The Subprime AI Crisis

We recorded 50 claims from this post. 15 were checked against evidence. 35 can’t be checked. 31 more are held. Read the post at its source.

Claims from this post (50)

In the order they appear in the post. Each quotation is the speaker’s words in this post. A label such as “Supported” or “Can’t be checked” is the claim’s current assessment; its page has the evidence and the reasons.

  1. Assessment: Can’t be checkedSaid

    “I also fear that said collapse could be ruinous to big tech, deeply damaging to the startup ecosystem, and will further sour public support for the tech industry.”

    What we tested: He fears that the collapse of the generative AI boom could be ruinous to big tech and deeply damaging to the startup ecosystem, and will further sour public support for the tech industry.

  2. Assessment: UnresolvedSaid

    “These are all signs that things are beginning to collapse.”

    What we tested: The o1 launch, rumored OpenAI price increases, Scale AI layoffs and OpenAI departures are signs the AI boom is beginning to collapse.

  3. Assessment: Can’t be checkedSaid

    “I’m going to paint you a bleak picture — not just for the big AI players, but for tech more widely, and for the people who work at tech companies — and tell you why I think the conclusion to this sordid saga, as brutal and damaging as it will be, is coming sooner than you think.”

    What we tested: The end of the generative AI boom will be brutal and damaging and is coming sooner than people think.

  4. Assessment: UnresolvedSaid

    “As I've explained in agonizing detail in the past, OpenAI will have to continue to raise more money than any startup has ever raised in history, in perpetuity, to survive.”

    What we tested: To survive, OpenAI will have to keep raising more money than any startup has ever raised, in perpetuity.

  5. Assessment: Can’t be checkedSaid

    “As I discussed at the end of July, OpenAI needs to raise at least $3 billion — but more like $10 billion to survive, as it is on course to lose $5 billion in 2024, a number that's likely to increase as more complex models demand more compute and more training data, with Anthropic CEO Dario Amodei predicting that future models may cost as much as $100 billion to train.”

    What we tested: OpenAI's annual losses are likely to increase beyond $5 billion as more complex models demand more compute and training data.

  6. Assessment: UnresolvedSaid

    “Given OpenAI’s existing profit-sharing deals with Microsoft — not to mention its deep, deep unprofitability — any such returns are, at the best, theoretical.”

    What we tested: Given Microsoft's profit-sharing deal and OpenAI's deep unprofitability, returns to OpenAI investors are at best theoretical.

  7. Assessment: Can’t be checkedSaid

    “In effect, investors aren't getting a piece of OpenAI, or any kind of control over OpenAI, but rather shares in the future profits of a company that loses over five billion dollars a year, and will likely lose more in 2025, if it makes it that far.”

    What we tested: OpenAI, which loses over $5 billion a year, will likely lose more in 2025, if it makes it that far.

  8. Assessment: Can’t be checkedSaid

    “While The Information reported that OpenAI's revenue is $3.5 to $4.5 billion a year in July, The New York Times reported last week that OpenAI's annual revenues have "now topped $2 billion," which would mean that the end-of-year numbers will likely trend toward the lower end of the estimate.”

    What we tested: OpenAI's 2024 end-of-year revenue will likely trend toward the lower end of the $3.5–4.5 billion estimate.

  9. Assessment: UnresolvedSaid

    “In short, OpenAI is burning money, will only burn more money, and to continue burning more money it will have to raise money from investors that are signing a document that says "we may never make a profit."”

    What we tested: OpenAI is burning money, will only burn more money, and to keep doing so will have to raise money from investors who sign a document saying it may never make a profit.

  10. Assessment: Can’t be checkedSaid

    “It has massive, intractable problems as a result of these models being probabilistic, meaning that they don't know anything, they're just generating an answer (or an image, or a translation, or a summary) based on training data, something that model developers are running out of at an incredible pace.”

    What we tested: Generative AI has intractable problems because its models are probabilistic and don't know anything, and developers are rapidly running out of training data.

  11. Assessment: Can’t be checkedSaid

    “And even if they did, it isn't clear whether generative AI actually provides much business value at all.”

    What we tested: It is not clear that generative AI provides much business value at all.

  12. Assessment: Can’t be checkedSaid

    “Last week, we got our first real, definitive glimpse of what’s around that corner that future. And boy, was it underwhelming.”

    What we tested: OpenAI's o1 launch last week, the first real, definitive glimpse of the next generation of models, was underwhelming.

  13. Assessment: Can’t be checkedSaid

    “While this sounds like a seismic breakthrough, or even another step towards the highly-feted artificial general intelligence (AGI) — it isn’t — and you can tell by the fact that OpenAI opted to release o1 as its own standalone product, rather than a number update to GPT.”

    What we tested: o1's chain-of-thought approach is not a seismic breakthrough or a step toward AGI.

  14. Assessment: UnresolvedSaid

    “This is because, despite there being a model that checks its work, the work-checking part of the model is still capable of hallucinations.”

    What we tested: o1 is more prone to hallucinations because its work-checking component can itself hallucinate.

  15. Assessment: Can’t be checkedSaid

    “It has described the reinforcement training process as “thinking” and “reasoning,” when, in fact, it’s making guesses, and then guessing on the correctness of those guesses at each step, where the end destination is often something that can be known in advance.”

    What we tested: o1 does not think or reason; it makes guesses and then guesses about the correctness of those guesses.

  16. Assessment: Can’t be checkedSaid

    “All of the things it generates to “consider” an answer (to be clear, this model is not “thinking”) are also charged for, making complex answers for things like coding likely incredibly expensive.”

    What we tested: Because hidden reasoning tokens are charged, complex o1 answers such as coding are likely very expensive.

  17. Assessment: Can’t be checkedSaid

    “OpenAI is shoving Strawberry out the door as a means of proving to investors (and the greater public) that the AI revolution is still here, and what they have is a clunky, unexciting and expensive model.”

    What we tested: OpenAI is pushing out o1 (Strawberry) to prove to investors and the wider public that the AI revolution is still here, but what it has is a clunky, unexciting and expensive model.

  18. Assessment: Can’t be checkedSaid

    “In fact, I'd argue that o1 shows that OpenAI is both desperate and out of ideas.”

    What we tested: o1 shows that OpenAI is desperate and out of ideas.

  19. Assessment: ContradictedSaid

    “The prices are not decreasing, the software is not becoming more useful, and the "next generation" model that we’ve been hearing about since last November has turned out to be a dud.”

    What we tested: Generative AI prices are not decreasing, the software is not becoming more useful, and OpenAI's next-generation model is a dud.

    What happened: OpenAI announced lower input and output token rates for the updated GPT-3.5 Turbo offering. This directly contradicts the broad claim that model prices were not decreasing. (OpenAI, )

  20. Assessment: UnresolvedSaid

    “If any of these lawsuits prevail, it would be calamitous for OpenAI and Anthropic, and even more so for Google and Meta, whose Gemini and Llama models use datasets including millions of artists' work, mostly because it's "virtually impossible" for AI models to forget training data, meaning that they'd need to be retrained from scratch, which would cost billions of dollars and dramatically reduce their efficacy at tasks they're already not particularly good at.”

    What we tested: If any of these copyright lawsuits prevail, it would be calamitous for OpenAI and Anthropic and even more so for Google and Meta, mostly because models can't practically forget training data and would need costly retraining from scratch that would cost billions and dramatically reduce their efficacy.

  21. Assessment: Can’t be checkedSaid

    “We’re not “in the early days.””

    What we tested: Generative AI is not in its early days, given over $150 billion spent since November 2022.

  22. Assessment: Can’t be checkedSaid

    “That it's artificial intelligence.”

    What we tested: It is a lie that generative AI is artificial intelligence.

  23. Assessment: Can’t be checkedSaid

    “That it will become artificial intelligence.”

    What we tested: It is a lie that generative AI will become artificial intelligence.

  24. Assessment: Can’t be checkedSaid

    “That it is inevitable.”

    What we tested: It is a lie that generative AI is inevitable.

  25. Assessment: UnresolvedSaid

    “Putting aside terms like "performance" — as they are largely used as a means of generating things "accurately" or "faster" rather than being good at stuff — Large Language Models have effectively plateaued.”

    What we tested: Setting aside 'performance' measures of accuracy or speed, Large Language Models have effectively plateaued.

  26. Assessment: ContradictedSaid

    “Nobody is trying to make these things more efficient, or at the very least nobody has succeeded in doing so.”

    What we tested: Nobody has succeeded in making LLMs more efficient.

    What happened: The developers report serving the existing LLM traffic with half as many GPUs after changing the serving software. This is a concrete resource-efficiency improvement before publication. (vLLM project, UC Berkeley researchers, )

  27. Assessment: UnresolvedSaid

    “My concern is that I believe we’re in the midst of a subprime AI crisis, where thousands of companies have integrated generative AI at prices that are far from stable, and even further from profitable.”

    What we tested: We are in a subprime AI crisis in which thousands of companies have integrated generative AI at unstable, unprofitable prices.

  28. Assessment: UnresolvedSaid

    “These models are built by two companies that are deeply unprofitable (Anthropic is on course to lose $2.7 billion this year), and that have pricing designed to get more customers rather than make any kind of profit.”

    What we tested: OpenAI and Anthropic are deeply unprofitable, with pricing designed to win customers rather than profit; Anthropic is on course to lose $2.7 billion in 2024.

  29. Assessment: Can’t be checkedSaid

    “Based on how unprofitable they are, I hypothesize that if OpenAI or Anthropic charged prices closer to their actual costs, there would be a ten-to-a-hundred-times increase in the price of API calls, though it's impossible to say how much without the actual numbers.”

    What we tested: He hypothesizes that if OpenAI or Anthropic charged prices closer to their actual costs, API prices would rise ten to a hundred times, though he says the amount is impossible to know without the actual numbers.

  30. Assessment: Can’t be checkedSaid

    “OpenAI is more than likely charging only a small percentage of what it likely costs to run its models, and can only continue to do so if it is able to continually raise more venture funding than has ever been raised and continue to receive preferential pricing from Microsoft, a company that recently mentioned that it considers OpenAI a competitor.”

    What we tested: OpenAI likely charges only a small fraction of its running costs and can only continue if it keeps raising record venture funding and receiving Microsoft's preferential pricing.

  31. Assessment: Can’t be checkedSaid

    “While we can’t know for sure, I imagine if this stuff was in any way profitable, they’d be talking about the revenue they were receiving from it.”

    What we tested: While admitting he can't know for sure, he imagines that if generative AI were in any way profitable for the hyperscalers, they would be talking about the revenue from it, and they are not.

  32. Assessment: UnresolvedSaid

    “At the beginning of August, Microsoft, Amazon, and Google all took a beating from the markets for their massive capital expenditures related to AI, and all three of them will face the wheel next quarter if they can't show a significant increase in revenue from the combined $150bn (if not more) that they've put into new data centers and NVIDIA GPUs.”

    What we tested: Microsoft, Amazon and Google will be punished by markets next quarter if they cannot show significant revenue increases from their AI capex.

  33. Assessment: Can’t be checkedSaid

    “What's important to remember here is that other than AI, big tech is out of ideas.”

    What we tested: Other than AI, big tech has no new growth ideas.

  34. Assessment: Can’t be checkedSaid

    “This is the most likely path for Amazon and Google, who, while desperate to make Wall Street happy, still have their own profitable monopolies to point to — for now, at least.”

    What we tested: The most likely path for Amazon and Google is a Meta-style 'year of efficiency': reducing capital expenditures, laying people off, and promising to slow investment somewhat, since they still have profitable monopolies to point to, for now.

  35. Assessment: Can’t be checkedSaid

    “I deeply worry about the prospect of mass layoffs being used to fund the movement, and nothing about the last few years makes me think they’ll do the right thing and walk away from AI.”

    What we tested: He deeply worries that mass layoffs may be used to fund big tech's AI push, and nothing about recent years makes him think they will do the right thing and walk away from AI.

  36. Assessment: Can’t be checkedSaid

    “I worry, somewhat, that option 2 is far more possible: that these companies are deeply committed to the idea that "AI is the future," and their cultures are so thoroughly disconnected from the creation of software that solves the problems that real people face that they'll burn the entire company to the ground.”

    What we tested: Big tech may well choose to cut costs and lay off workers to sustain generative AI spending, burning their companies to the ground.

  37. Assessment: Can’t be checkedSaid

    “All this could’ve been stopped earlier if the media had actually held them accountable.”

    What we tested: The generative AI hype could have been stopped earlier if the media had held tech companies accountable.

  38. Assessment: SupportedSaid

    “I believe that, at the very least, Microsoft will begin reducing costs in other areas of its business as a means of helping sustain the AI boom.”

    What we tested: Microsoft will begin reducing costs in other parts of its business to help sustain its AI spending.

    What happened: Reporting based on an interview with Microsoft president Brad Smith links operating-cost pressure to the infrastructure investment discussed in the article, alongside implemented workforce cuts. Together with the Gaming memo, this supports AI investment as a contributing reason for reductions elsewhere. (GeekWire, )

  39. Assessment: Can’t be checkedSaid

    “Based on The Information's estimates, Microsoft has somewhere between 400,000 and 4 million users for its Office Copilot features, meaning that there's a decent chance that Microsoft has built out capacity that isn't getting used.”

    What we tested: There is a decent chance Microsoft has built data center capacity for Copilot that is not being used.

  40. Assessment: Can’t be checkedSaid

    “I realize I sound a little crazy suggesting this, but back in March I made the point that I could find no companies that had integrated generative AI in a way that has truly benefited their bottom line, and just under six months later, I'm still looking.”

    What we tested: He has found no company whose generative AI integration has truly benefited its bottom line.

  41. Assessment: Can’t be checkedSaid

    “While there may be companies "integrating AI" that are driving some degree of spend on Microsoft Azure, Amazon Web Services and Google Cloud, I hypothesize a lot of this demand is driven by investor sentiment, with companies “investing in AI” to keep the markets happy rather than any cost/benefit analysis or actual utility.”

    What we tested: He hypothesizes that a lot of cloud demand from companies integrating AI is driven by investor sentiment, with companies investing in AI to keep markets happy rather than for cost/benefit or actual utility.

  42. Assessment: Can’t be checkedSaid

    “After developing and launching these features, they find customers don't want to pay for them, as Microsoft has found with its 365 Copilot, and if they can't find a way to make them pay for it now, during the AI boom, they're going to be hard pressed to once people's bosses stop asking them to "get in on AI."”

    What we tested: In one of several hypothesized scenarios, companies that can't get customers to pay for AI features now, during the boom, will be hard pressed to once bosses stop pushing people to get in on AI.

  43. Assessment: Can’t be checkedSaid

    “This point also illustrates one of the biggest problems with generative AI — that, while "powerful" in some vague way, said power mostly boils down to "generates stuff based on other stuff" rather than any "intelligence."”

    What we tested: Generative AI's power mostly amounts to generating stuff based on other stuff, not intelligence.

  44. Assessment: Can’t be checkedSaid

    “I believe that a lot of businesses are "trying" AI at the moment, and once those trials end (Gartner predicts that 30% of generative AI projects will be abandoned after their proof of concepts by end of 2025), they'll likely stop paying for the extra features, or stop integrating generative AI into their companies' products.”

    What we tested: Once their AI trials end, many businesses will likely stop paying for AI features or stop integrating generative AI into products.

  45. Assessment: UnresolvedSaid

    “Even though Big Tech could keep funding this boom — after all, they’re pretty much entirely responsible for its existence — that won’t help the startups that won't be able to afford to run their companies, after getting used to discounted rates.”

    What we tested: Even if Big Tech keeps funding the boom, that won't help the startups, used to discounted rates, that won't be able to afford to run their companies when prices rise.

  46. Assessment: Can’t be checkedSaid

    “In any case, it's time to accept that the money isn't there.”

    What we tested: People should accept that the money in generative AI isn't there.

  47. Assessment: Can’t be checkedSaid

    “Yet, unlike cryptocurrency and the metaverse, everybody has joined the party and decided to burn money pursuing an unsustainable, unreliable, unprofitable, and environmentally-destructive boondoggle sold to customers and businesses as "artificial intelligence" that will "automate everything" without it ever having a path to do so.”

    What we tested: Unlike with crypto and the metaverse, everybody has joined in burning money on generative AI, an unsustainable, unreliable, unprofitable, environmentally destructive boondoggle sold as AI that will automate everything without any path to doing so.

  48. Assessment: Can’t be checkedSaid

    “I hypothesize a kind of subprime AI crisis is brewing, where almost the entire tech industry has bought in on a technology sold at a vastly-discounted rate, heavily-centralized and subsidized by big tech.”

    What we tested: He hypothesizes that a kind of subprime AI crisis is brewing, with almost the entire tech industry bought in on a technology sold at vastly discounted rates, heavily centralized and subsidized by big tech.

  49. Assessment: UnresolvedSaid

    “At some point, the incredible, toxic burn-rate of generative AI is going to catch up with them, which in turn will lead to price increases, or companies releasing new products and features with wildly onerous rates — like the egregious $2-a-conversation rate for Salesforce’s “Agentforce” product — that will make even stalwart enterprise customers with budget to burn unable to justify the expense.”

    What we tested: Generative AI's burn rate will eventually force price increases or onerous pricing that even well-funded enterprise customers cannot justify.

  50. Assessment: Can’t be checkedSaid

    “But the tech industry is building toward a grotesque reckoning with a lack of creativity enabled by an economy that rewards growth over innovation, monopolization over loyalty, and management over those who actually build things.”

    What we tested: The tech industry is building toward a grotesque reckoning with a lack of creativity, enabled by an economy that rewards growth over innovation, monopolization over loyalty, and management over builders.

Held from this post (31)

These records did not pass every check, so nothing about them is published. Held is not a finding.