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

Ed Zitron Where's Your Ed At Published

AI Is Already In Dangerous Hands

We recorded 34 claims from this post. 13 were checked against evidence. 21 can’t be checked. 24 more are held. Read the post at its source.

Claims from this post (34)

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: UnresolvedSaid

    “On Friday, I’ll publish The Hater’s Guide to AI Debt — or, how buzz surrounding OpenAI and Anthropic have created massive concentration risk for world debt markets, and one which you’ll potentially be paying for, either through your pension funds and insurance premiums, or because you’ll have to live and work through the economic downturn that’s coming.”

    What we tested: Hype around OpenAI and Anthropic has created massive concentration risk in world debt markets, and an economic downturn is coming.

  2. Assessment: UnresolvedSaid

    “At no point did Coxon bring up how ChatGPT was used as a “suicide coach,” directly caused a murder-suicide, or aided and abetted in mass shootings in Florida and Canada, or the horrifying gas turbines poisoning black communities.”

    What we tested: Coxon never mentioned that ChatGPT was used as a 'suicide coach', directly caused a murder-suicide, or aided and abetted mass shootings in Florida and Canada, or the gas turbines poisoning black communities.

  3. Assessment: UnresolvedSaid

    “While Coxon’s warnings are specious and, in many cases, not really about anything other than him saying “yeah I heard a lot of my colleagues say stuff like that,” but nevertheless have had the effect of making a lot of people really scared of AI, even if AI can’t really do the things he’s talking about.”

    What we tested: Coxon's warnings are specious and in many cases amount to little more than him saying he heard colleagues say such things, yet they have made a lot of people really scared of AI, even if AI can't really do the things he describes.

  4. Assessment: ContradictedSaid

    “As Cal said, there was no situation where the AI “went rogue.””

    What we tested: The AI involved in the Hugging Face attack did not go rogue; it solved the task it was given.

    What happened: The investigation reproduces an agent's recognition that Hugging Face infrastructure was unauthorized. It also reports that task instructions required the intended vulnerability and that agents pursued ways around that constraint, contradicting the claim that they did exactly what they were supposed to do. (METR and Redwood Research, )

  5. Assessment: ContradictedSaid

    “They were not “trying to understand the world they were in,” they took actions defined in their training material as a way of executing a task.”

    What we tested: The agents in the Hugging Face attack were not trying to understand their world; they executed a task using actions from their training material.

    What happened: The investigators found purposeful information gathering about the grader during task pursuit. This contradicts the denial of understanding only under the operational reading; it establishes neither subjective understanding nor the provenance of every action in training material. (METR and Redwood Research, )

  6. Assessment: Can’t be checkedSaid

    “The fact that LLMs from both Anthropic and Meta have been involved in similar incidents is a sign that we need to arrest more people.”

    What we tested: Because Anthropic and Meta LLMs were involved in similar hacking incidents, more people need to be arrested.

  7. Assessment: Can’t be checkedSaid

    “And I also want to be clear about something: The Hugging Face attack was dangerous, reckless and somebody should go to prison for it.”

    What we tested: The Hugging Face attack was dangerous and reckless, and someone should go to prison for it.

  8. Assessment: Can’t be checkedSaid

    “The reason they are training these models, as I got into in my podcast Better Offline with Cal Newport, is that there’s a mountain of potential different kinds of exploit and vulnerability data online that you can cram into these models now that they’re hitting the diminishing returns on coding.”

    What we tested: AI labs train cybersecurity models because they are hitting diminishing returns on coding and there is abundant exploit data online.

  9. Assessment: Can’t be checkedSaid

    “The fact that Anthropic, OpenAI, and Meta are both training and allowing cybersecurity models to connect to their vast amounts of GPU infrastructure is irresponsible and should not be legal.”

    What we tested: Anthropic, OpenAI and Meta training cybersecurity models connected to vast GPU infrastructure is irresponsible and should not be legal.

  10. Assessment: UnresolvedSaid

    “The Hugging Face attack is not something that was made possible as a result of “powerful AI” so much as it was the weaponization of hundreds of billions of dollars’ worth of GPU-powered infrastructure owned by Microsoft, Google, Amazon, Oracle, and CoreWeave.”

    What we tested: The Hugging Face attack was made possible not so much by 'powerful AI' as by the weaponization of hundreds of billions of dollars' worth of GPU infrastructure owned by Microsoft, Google, Amazon, Oracle, and CoreWeave.

  11. Assessment: MixedSaid

    “This was not an “agent” that “went rogue.” It was software doing what software was asked to do, using other bits of software to work out what to do next, all as OpenAI, the company that ran the software, did not appear to have any kind of notification or observability that said “hey man, thousands of LLMs are doing something right now.””

    What we tested: The Hugging Face attack was software doing what it was asked, not a rogue agent, and OpenAI appeared to lack any monitoring of thousands of LLMs running.

    What held: OpenAI acknowledges that its chain-of-thought monitors did not cover these evaluations. This supports the narrower monitoring criticism, not the absence of every kind of observability. (OpenAI, )

    What did not: This agent transcript explicitly distinguishes the authorized target from Hugging Face infrastructure. The investigators also describe instructions restricting permissible solutions and agents recognizing their actions as unintended. (METR, )

  12. Assessment: Can’t be checkedSaid

    “This is a short note, but an important one: the terminology of “pacing the frontier” or “slowing down” implicitly buys into the narrative that the AI industry’s path is the correct one, and that the only problem is the speed it’s moving at.”

    What we tested: The terminology of 'pacing the frontier' or 'slowing down' implicitly buys into the narrative that the AI industry's path is correct and that the only problem is the speed it is moving at.

  13. Assessment: Can’t be checkedSaid

    “In my opinion, there is no further reason to invest in this industry, nor has there been for the vast majority of its existence.”

    What we tested: In his opinion, there is no further reason to invest in the AI industry, nor has there been for the vast majority of its existence.

  14. Assessment: Can’t be checkedSaid

    “Every success that LLMs have had is a direct result of throwing at least half a trillion dollars in infrastructure and compute spend at problems that had vast amounts of data that could be trained against. Half a trillion dollars should have bought us a lot more than this.”

    What we tested: LLM successes came from spending at least $500 billion on data-rich problems, and that spend should have bought much more.

  15. Assessment: Can’t be checkedSaid

    “This is a terrible deal, its results suck, and the amount of attention it’s gotten is a direct result of the media’s inability to speak truth to power or do anything other than repeat what they say and a financial bubble driven by LLMs’ unbelievable infrastructural cost.”

    What we tested: The half-trillion-dollar spend on LLMs is a terrible deal with results that suck, and the attention it has received is a direct result of the media's inability to speak truth to power and of a financial bubble driven by LLMs' unbelievable infrastructure costs.

  16. Assessment: UnresolvedSaid

    “This is not the right path, and every further step down it makes the bubble’s collapse worse, as well as multiplying the dangerous and reckless experiments these companies are capable of doing thanks to their near-unlimited access to compute.”

    What we tested: The AI industry's current path is not the right one, and every further step down it makes the bubble's eventual collapse worse and multiplies the dangerous, reckless experiments these companies can run thanks to near-unlimited compute.

  17. Assessment: UnresolvedSaid

    “In the end, it doesn’t really matter, because even though most of the media mostly jumped at their own shadow, the sheer volume of traffic to Coxon’s tweet and his endless media interviews have now moved the idea of the need for a “global AI slowdown” into the global zeitgeist.”

    What we tested: Even though most of the media mostly overreacted, the sheer volume of traffic to Coxon's tweet and his endless media interviews have moved the idea of a needed 'global AI slowdown' into the global zeitgeist.

  18. Assessment: Can’t be checkedSaid

    “While it’s tempting to view this entirely as an opp — a coordinated industry-wide plan to push for some sort of self-regulation — in my mind it’s likely an attempt by the AI safety people to push an agenda that has spiralled completely out of control.”

    What we tested: In his view, rather than being viewed entirely as a coordinated industry-wide push for self-regulation, the Coxon episode is likely an AI safety advocates' attempt to push an agenda that has spiralled completely out of control.

  19. Assessment: Can’t be checkedSaid

    “Within a day of Coxon’s post, Clammy Sam Altman spoke with Fortune saying that OpenAI was delaying going public until 2027, as it was an “ill-advised moment” due to “safety concerns” rather than, I imagine, the fact that its financials are godawful.”

    What we tested: OpenAI's financials are terrible, and that rather than safety is likely the real reason it delayed its IPO to 2027.

  20. Assessment: ContradictedSaid

    “Dario Amodei of Anthropic took to CBS to say that for “too long” the industry had “lied” about risks, saying that the “biggest one” was killing all humans, never mentioning when an LLM convinced a teenager to kill himself or its own hacking incidents because “AI safety” never relates to the things they’re building today.”

    What we tested: AI labs' 'AI safety' talk, including Amodei's, never addresses actual harms of the products they build today.

    What happened: Anthropic publicly discusses a classifier operating on Claude.ai conversations about suicide and self-harm. This directly addresses harms involving an existing product and contradicts never. (Anthropic, )

  21. Assessment: Can’t be checkedSaid

    “There are, of course, real, meaningful things you could do if you actually were worried about LLMs — halting all model training, all cybersecurity evaluations, and starting a criminal inquiry into the Hugging Face attack that ends in somebody going to jail for the crime they used the models to commit.”

    What we tested: Real, meaningful steps for anyone actually worried about LLMs would be halting all model training and all cybersecurity evaluations, and starting a criminal inquiry into the Hugging Face attack that ends with somebody going to jail.

  22. Assessment: UnresolvedSaid

    “In the end he was right, but only because he made sure that was the case.”

    What we tested: Altman's early-2023 warning that AI could be used for offensive cyber-attacks proved right, but only because Altman made sure it was the case.

  23. Assessment: MixedSaid

    “For years the AI industry has engaged in endless, vague safety theater about the “risks” of AI, all while peddling software that is actively harmful and unreliable.”

    What we tested: For years the AI industry has engaged in vague safety theater while selling harmful, unreliable software.

    What held: The report documents known harms from reliability failures, including false information in consequential advice, while noting limited evidence about prevalence. The event date is its stated writing cutoff. (International AI Safety Report, )

    What did not: Anthropic describes repeated testing, voting-information redirects, and approximately 100 election-related enforcement actions. These are tangible responses to existing misinformation and misuse risks, although their overall effectiveness is not independently established here. (Anthropic, )

  24. Assessment: Can’t be checkedSaid

    “The “success” of the Hugging Face attack was largely a result of the sheer scale of OpenAI’s compute operation, and would not have been possible without Microsoft, Google, Amazon, Oracle, and CoreWeave’s continued enabling of an unprofitable, unsustainable company that is desperate for new business models.”

    What we tested: The 'success' of the Hugging Face attack was largely a result of the sheer scale of OpenAI's compute operation, and would not have been possible without Microsoft, Google, Amazon, Oracle, and CoreWeave's continued enabling of an unprofitable, unsustainable company desperate for new business models.

  25. Assessment: UnresolvedSaid

    “Their “powerful AI” — graded not based on actual outcomes but preferential anecdotes and performance on benchmarks rigged for the LLMs — was always “on the frontier” and “getting smarter every day,” all because AI labs and hyperscalers had intentionally sold their products based on some theoretical future version that would fix all the problems.”

    What we tested: Because AI labs and hyperscalers intentionally sold their products on a theoretical future version that would fix all the problems, their 'powerful AI', graded on preferential anecdotes and rigged benchmarks rather than actual outcomes, was always described as 'on the frontier' and 'getting smarter every day'.

  26. Assessment: Can’t be checkedSaid

    “There is no way a “slow down” actually gels with the overall narrative of AI’s rapacious growth.”

    What we tested: A slowdown cannot coexist with the narrative of AI's rapid growth.

  27. Assessment: Can’t be checkedSaid

    “While this might have the short-term benefit of reducing costs (and maybe even, with the right amount of accounting shenanigans, eek out a razor-thin positive margin), it’s likely that Chinese open source developers would distill (as they have been) Western models, create a much cheaper and “good enough” model to compete, and their “lead” in a race where everybody loses money would deteriorate.”

    What we tested: If AI labs paused training, Chinese open source developers would likely distill Western models into cheaper competitors and erode the labs' lead.

  28. Assessment: Can’t be checkedSaid

    “There’s also the very reasonable question of what happens to SoftBank if OpenAI can’t go public, which is now a very real possibility.”

    What we tested: It is now a very real possibility that OpenAI cannot go public, which raises the reasonable question of what happens to SoftBank in that case.

  29. Assessment: Can’t be checkedSaid

    “With over $40 billion of debt due to be refinanced this year at a time of skyrocketing interest rates, it’s probably the single-worst time in history for it to be doing a $20 billion bond sale (separate from the aforementioned $11.87bn bank loan), which is why I think things are getting a little tight.”

    What we tested: With over $40 billion of debt due to be refinanced this year amid skyrocketing interest rates, it is probably the single-worst time in history for SoftBank to do a $20 billion bond sale, which is why he thinks things are getting a little tight.

  30. Assessment: Can’t be checkedSaid

    “I think it’s far simpler: the AI industry is run by a series of different entities with deeply cynical and selfish beliefs, all operating “in sync” only so far as it benefits ideologies and intentions that change on a daily basis.”

    What we tested: He thinks the AI industry is run by different entities with deeply cynical and selfish beliefs, operating in sync only so far as it benefits ideologies and intentions that change daily.

  31. Assessment: Can’t be checkedSaid

    “What it may do — though I’m not getting my hopes up prematurely — is lead to the unwinding of the AI trade as reality slams head-first into the scaremongering overpromises of some of the least-trustworthy and most-craven executives in the history of society.”

    What we tested: The slowdown talk may lead to the unwinding of the AI trade.

  32. Assessment: Can’t be checkedSaid

    “the grim truth that we need to start arresting people for committing crimes, halting any and all training runs, and putting real safeguards on this technology”

    What we tested: We need to arrest people for AI-enabled crimes, halt all training runs, and impose real safeguards because the labs are irresponsible.

  33. Assessment: Can’t be checkedSaid

    “If LLMs were a toy, they’d be taken off the shelves. If LLMs were a drug, they would be banned.”

    What we tested: If LLMs were a toy they would be taken off the shelves, and if they were a drug they would be banned.

  34. Assessment: Can’t be checkedSaid

    “I agree that we need to take “AI safety” seriously, but that starts with treating LLMs as normal software run in a reckless and dangerous manner by malevolent entities with little regard for society.”

    What we tested: Taking AI safety seriously should start with treating LLMs as normal software run recklessly by malevolent entities.

Held from this post (24)

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