Supported claims, page 1 of 6
Claims by Gary Marcus and Ed Zitron. Deadlines are judged as of 2026-10-29.
Supported claims (531)
The dated evidence matches what the claim committed to, under the criteria frozen before grading.
Page 1 of 6: claims 1 to 100 of 531, ordered by the date first said, newest first. To find a claim by its words, use search.
Assessment: SupportedSaid
Gary Marcus Observation
“To say nothing of the months of attacks from OpenAI and others on places like Hugging Face and the Australian government.”
Our reading: For months there have been attacks from OpenAI and others on places such as Hugging Face and the Australian government.
No horizon Evidence as of
Limitation: A stronger reading requiring months of attacks against each named target remains unproven.
Source: Breaking news: U.S. banks are taking hits from AI agents
Assessment: SupportedSaid
Gary Marcus Observation
“There is also a growing, ongoing wave of cyberattacks in Japan.”
Our reading: There is a growing, ongoing wave of cyberattacks in Japan.
No horizon Evidence as of
Limitation: The quantitative series covers selected publicly disclosed web breaches.
Source: Breaking news: U.S. banks are taking hits from AI agents
Assessment: SupportedSaid
Gary Marcus Causal claim
“Current models are already causing serious problems across society, in domains ranging from education to politics (via deepfakes and wholesale disinformation) to banking.”
Our reading: Current AI models are already causing serious problems across society, in domains ranging from education to politics (via deepfakes and wholesale disinformation) to banking.
No horizon Evidence as of
Limitation: Support depends on treating current models as existing generative AI technology rather than only the latest releases.
Source: Breaking news: U.S. banks are taking hits from AI agents
Assessment: SupportedSaid
Gary Marcus Observation
“P.S. In other news, early generative AI darling Character.AI has filed for bankruptcy, and Coreweave bonds have reached 15%, suggesting that credit markets have had enough.”
Our reading: Character.AI has filed for bankruptcy.
No horizon Evidence as of
Limitation: This assessment establishes the filing, not liquidation, shutdown, or the eventual result of the bankruptcy.
Assessment: SupportedSaid
Gary Marcus Observation
“Chatbots have been implicated in a number of suicides, and even mass murder.”
Our reading: Chatbots have been implicated in a number of suicides and even in mass murder.
No horizon Evidence as of
Limitation: The allegations are not findings of causation or liability.
Assessment: SupportedSaid
Ed Zitron Observation
“In 2025, Anthropic lost over $8 billion on $4.6 billion in revenue.”
Our reading: Anthropic's revenue for calendar 2025 was about $4.6 billion.
No horizon Evidence as of
Limitation: Zitron's source was a Reuters report on a leaked draft. The record checks his figure against the S-1 as filed.
Source: Dead Money
Assessment: SupportedSaid
Gary Marcus Observation
“It didn’t just attack Australia’s government servers.”
Our reading: OpenAI's software attacked Australia's government servers, among other targets.
No horizon Evidence as of
Limitation: Only one Australian government system had confirmed unauthorised access, so the plural 'servers' is only partly borne out.
Source: BREAKING: OpenAI’s epic meltdown — and how Sam Altman is blowing up Jensen Huang’s reputation
Assessment: SupportedSaid
Gary Marcus Observation
“Even in that long tweet above, they didn’t even dare report how many incidents there have been.”
Our reading: OpenAI's long tweet did not report how many incidents there had been.
No horizon Evidence as of
Limitation: This finding concerns the post itself, not whether counts appeared elsewhere.
Source: BREAKING: OpenAI’s epic meltdown — and how Sam Altman is blowing up Jensen Huang’s reputation
Assessment: SupportedSaid
Gary Marcus Observation
“Back in reality, the evidence that OpenAI is building something the company isn’t managing to control is vast and growing.”
Our reading: There is vast and growing evidence that OpenAI is building systems it is not managing to control.
No horizon Evidence as of
Limitation: The September disclosure adds evidence about the July attack; it does not establish increasing incident frequency [e1, e2].
Source: BREAKING: OpenAI’s epic meltdown — and how Sam Altman is blowing up Jensen Huang’s reputation
Assessment: SupportedSaid
Gary Marcus Observation
“OpenAI’s software didn’t just attack Hugging Face.”
Our reading: OpenAI's software attacked Hugging Face, and its attacks went beyond Hugging Face.
No horizon Evidence as of
Limitation: OpenAI describes some incidents as 'interactions' rather than attacks. The evidence does not settle whether each notified case was an intrusion.
Source: BREAKING: OpenAI’s epic meltdown — and how Sam Altman is blowing up Jensen Huang’s reputation
Assessment: SupportedSaid
Gary Marcus Observation
“Worse, in disclosing what happened, OpenAI has dragged their feet every along the way.”
Our reading: OpenAI has been slow at every step in disclosing what happened in the misaligned models incident.
No horizon Evidence as of
Limitation: 'Every step along the way' cannot be checked literally. It is read as a repeated pattern.
Source: BREAKING: OpenAI’s epic meltdown — and how Sam Altman is blowing up Jensen Huang’s reputation
Assessment: SupportedSaid
Gary Marcus Observation
“China is practically begging to collaborate.”
Our reading: China is eager to collaborate with the US on AI, contrary to claims it would never want to.
No horizon Evidence as of
Limitation: Support depends on reading the phrase as an emphatic description of public overtures.
Source: Dear President Trump, here’s a deal you can actually make that will make you look great
Assessment: SupportedSaid
Gary Marcus Observation
“Fast-forward to today, and basically Meta Muse is trying to do exactly what Facebook M tried to do, restaurant reservations, theatre tickets and so on.”
Our reading: Meta Muse is basically trying to do exactly what Facebook M tried to do, such as restaurant reservations and theatre tickets.
No horizon Evidence as of
Limitation: Support depends on reading basically as substantial task overlap and theatre tickets as an example of event booking.
Source: The secret behind Meta’s Muse
Assessment: SupportedSaid
Gary Marcus Observation
“No more than 30% of requests served by AI!”
Our reading: No more than 30% of Facebook M's requests were handled by AI.
No horizon Evidence as of
Limitation: The numerical evidence rests on one unnamed source's estimate.
Source: The secret behind Meta’s Muse
Assessment: SupportedSaid
Gary Marcus Observation
“Until now, I always assumed the game plan was to get a bailout if and when the economics fall apart, but at this point the big tech companies might be more screwed than I thought, given how toxic data centers have become.”
Our reading: AI data centers have become politically toxic.
No horizon Evidence as of
Limitation: The polling measures nearby data-center siting, not acceptance of all AI infrastructure nationally [e1].
Source: Data center madness
Assessment: SupportedSaid
Gary Marcus Observation
“Trump himself hasn’t entirely bolted — yet — but he’s certainly changed course sharply, moving from full anti AI regulation to strongly encouraging the voluntary preflight testing of frontier AI models.”
Our reading: Trump has sharply shifted from full opposition to AI regulation to strongly encouraging voluntary preflight testing of frontier AI models, though he has not entirely bolted yet.
No horizon Evidence as of
Limitation: 'Strongly encouraging' and 'full anti AI regulation' are characterizations. The verdict relies on a plain reading of them.
Source: Breaking: The Republican party is panicking over its ties to Big Tech
Assessment: SupportedSaid
Gary Marcus Observation
“Google has great distribution, through Android, Mail, Search, Docs, YouTube etc.”
Our reading: Google has strong distribution through Android, Mail, Search, Docs and YouTube.
No horizon Evidence as of
Limitation: The evidence is company-reported.
Assessment: SupportedSaid
Gary Marcus Observation
“Unfortunately, beyond that OpenAI does not report which problems they tried and failed at. They have given us a numerator without a denominator, always a worrisome sign.”
Our reading: OpenAI did not report which problems Astra tried and failed at, giving a numerator without a denominator.
No horizon Evidence as of
Limitation: This assessment covers the inspected public disclosures through August 3, 2026, not internal records.
Assessment: SupportedSaid
Gary Marcus Observation
“How good Astra is in open-ended real-world problems very much remains to be seen.”
Our reading: How good Astra is on open-ended real-world problems very much remains to be seen.
No horizon Evidence as of
Limitation: This assesses what public evidence established at publication, not whether Astra was capable of practical work.
Assessment: SupportedSaid
Gary Marcus Causal claim
“Math lends itself to two things: verification (using symbolic tools), and massive amounts of cheaply produced synthetic data where you can guarantee that the answers are correct. The same applies to coding — but it is not true in general.”
Our reading: Models shine at math and coding because those domains allow verification and cheap correct synthetic data, which is not true of domains in general.
No horizon Evidence as of
Limitation: The result depends on reading correctness relative to formal specifications.
Source: OpenAI’s amazing — but vastly oversold — new model Astra
Assessment: SupportedSaid
Gary Marcus Observation
“MIT Technology Review reported new, disturbing adventures in jailbreaking”
Our reading: MIT Technology Review reported new jailbreaking adventures, which Marcus calls disturbing.
No horizon Evidence as of
Limitation: This assessment verifies the publication and its subject, not the article's broader scientific conclusions.
Source: The seven most shambolic things that happened in AI today.
Assessment: SupportedSaid
Gary Marcus Observation
“We are clearly not there yet, either.”
Our reading: Humans have not merged with AI via brain-computer interfaces to become much smarter, so Kurzweil's singularity has not arrived.
No horizon Evidence as of
Limitation: Much smarter has no specified threshold.
Source: Sorry, Sam and Elon, we have not reached the Singularity
Assessment: SupportedSaid
Gary Marcus Observation
“The idea that we have reached that point is laughable.”
Our reading: Current AI has not reached a level beyond AGI that can do anything an expert human can and much more.
No horizon Evidence as of
Limitation: ARC-AGI-3 measures a particular kind of adaptive reasoning, not intelligence as a whole.
Source: Sorry, Sam and Elon, we have not reached the Singularity
Assessment: SupportedSaid
Gary Marcus Observation
“In the intervening nine months, the market has grown wary of all this circular financing stuff.”
Our reading: Between September 2025 and July 2026 the market grew wary of AI circular financing deals.
No horizon Evidence as of
Limitation: The evidence compares different announcements and companies.
Assessment: SupportedSaid
Gary Marcus Observation
“that models will continue to get more efficient and less expensive, but that hallucinations and reliability problems will persist.”
Our reading: Models have gotten more efficient and cheaper, but hallucinations and reliability problems have persisted.
No horizon Evidence as of
Limitation: The concrete within-period efficiency comparison is vendor-reported.
Assessment: SupportedSaid
Gary Marcus Causal claim
“But the Mythos moment got The White House to consider and in fact implement preflight testing to some degree (though not with the transparency or independence or breadth or mandatoriness that I believe to be critical).”
Our reading: The Mythos moment caused the White House to implement AI preflight testing to some degree, though without the transparency, independence, breadth or mandatoriness Marcus considers critical.
No horizon Evidence as of
Limitation: The causal link rests on news reporting. The order's text reportedly does not name Mythos.
Source: Breaking: Demis Hassabis endorses preflight safety testing for AI
Assessment: SupportedSaid
Gary Marcus Observation
“In my 2023 Senate testimony, in a 2023 essay here, and in my 2024 book Taming Silicon Valley, I strongly advocated for a system of preflight testing for AI models deployed at large scale, ideally mandatory, transparent and independent, modeled perhaps on the FDA’s analysis of costs and benefits for pharmaceuticals.”
Our reading: In his 2023 Senate testimony, a 2023 essay, and his 2024 book Taming Silicon Valley, Marcus strongly advocated preflight testing for AI models deployed at large scale, ideally mandatory, transparent and independent, perhaps modeled on the FDA's cost-benefit analysis for pharmaceuticals.
No horizon Evidence as of
Limitation: The evidence does not establish that every work separately included every safeguard or the exact FDA cost-benefit formulation.
Source: Breaking: Demis Hassabis endorses preflight safety testing for AI
Assessment: SupportedSaid
Gary Marcus Observation
“• Chinese models are gaining traction, fast.”
Our reading: Chinese AI models are rapidly gaining market traction.
No horizon Evidence as of
Limitation: The observations do not cover the full June 1–25 target period or establish continuous growth within it.
Assessment: SupportedSaid
Gary Marcus Observation
“In my view, the world has indeed gone in this direction. LLMs have become commodities, prices wars have been significant, profits for the LLM providers have been elusive.”
Our reading: Price wars among LLM providers have been significant.
No horizon Evidence as of
Limitation: The clearest evidence of directly responsive price cutting concerns China.
Source: The Generative AI Fizzle™
Assessment: SupportedSaid
Gary Marcus Observation
“Small wonder tokenmaxxing is dying;”
Our reading: Tokenmaxxing (maximizing AI token use) is in decline.
No horizon Evidence as of
Limitation: This is qualitative evidence of a corporate trend, not a representative prevalence estimate.
Source: Accenture: Then and now, and how it may signify things to come
Assessment: SupportedSaid
Gary Marcus Observation
“It took longer than I expected, but well, here we are:”
Our reading: The politics and inadequacy of generative AI guardrails, which Marcus predicted in 2024 would become central, has now become a central issue, though later than he expected.
No horizon Evidence as of
Limitation: Prominence is qualitative.
Assessment: SupportedSaid
Gary Marcus Observation
“Importantly, this is not an Anthropic problem, it’s a Generative AI problem.”
Our reading: The inability to block all jailbreaks is a problem with generative AI in general, not one specific to Anthropic.
No horizon Evidence as of
Limitation: The assessment was published in May, but the accessible source does not give individual experiment dates.
Assessment: SupportedSaid
Gary Marcus Observation
“The White House has acknowledged that some models might be risky; now we are just arguing about the details.”
Our reading: The White House has acknowledged some AI models might be risky, so the debate is now only about details of regulation.
No horizon Evidence as of
Limitation: Sacks’s explanation establishes the administration’s stated position, not the model’s actual danger.
Source: What Washington must do
Assessment: SupportedSaid
Gary Marcus Observation
“The ridiculous but fashionable idea of zero-regulation around AI, made popular by folks like Marc Andreessen, is (rightly) dead in the water.”
Our reading: The idea of zero regulation of AI is now, rightly, dead.
No horizon Evidence as of
Limitation: The result depends on the surrounding sentences narrowing the rhetoric to the White House’s acceptance of some model restrictions.
Source: What Washington must do
Assessment: SupportedSaid
Gary Marcus Observation
“The Commerce department just ordered an export control directive (quoting Anthropic) “to suspend all access to Fable 5 and Mythos 5” by foreign nationals — even their own employees.”
Our reading: The US Commerce Department issued an export control directive requiring Anthropic to suspend all access by foreign nationals, including its own employees, to Fable 5 and Mythos 5.
No horizon Evidence as of
Limitation: The directive itself was not found in a public government record. The evidence is Anthropic's account and Fortune's reporting.
Source: Breaking news: US Commerce Department effectively shuts down Anthropic’s latest models
Assessment: SupportedSaid
Gary Marcus Causal claim
“Now let’s consider last night’s Commerce Department cut-off-your-nose-to-spite-your-face export control order, which has basically led to the (temporary?) shutdown of Fable and Mythos, perhaps sparking a crisis in the US AI industry.”
Our reading: The Commerce Department's June 2026 export control order has basically led to the possibly temporary shutdown of Anthropic's Fable and Mythos.
No horizon Evidence as of
Limitation: The evidence establishes suspended customer access, not cessation of all internal model activity.
Assessment: SupportedSaid
Gary Marcus Observation
“His company did however recently support state law in Illinois that “would shield AI labs from liability in cases where AI models are used to cause serious societal harms, such as death or serious injury of 100 or more people or at least $1 billion in property damage”– pretty much the opposite of what he said to Senator Durbin in his deceitful, lie-filled testimony.”
Our reading: OpenAI recently supported an Illinois state law that would shield AI labs from liability in cases where AI models are used to cause serious societal harms, which Marcus calls pretty much the opposite of what Altman told Senator Durbin in his 2023 testimony.
No horizon Evidence as of
Limitation: This outcome applies to the legislative-support proposition, not the allegation of deceit.
Source: Maybe Section 230 doesn’t shield AI companies from liability, after all
Assessment: SupportedSaid
Gary Marcus Observation
“And OpenAI still has yet to turn a profit, and it’s still not clear they ever will.”
Our reading: OpenAI has not yet turned a profit and it is unclear whether it ever will.
No horizon Evidence as of
Limitation: The figures come from press reports of shareholder documents, not audited public filings. Reuters said it could not immediately verify the report.
Assessment: SupportedSaid
Gary Marcus Observation
“Anthropic has taken a lot of the business market and OpenAI has lost a lot of it.”
Our reading: Anthropic has taken a large share of the business market while OpenAI has lost much of it.
No horizon Evidence as of
Limitation: Enterprise LLM API share does not measure the entire business AI market [e1].
Assessment: SupportedSaid
Gary Marcus Causal claim
“And the reason that those tend to be more expensive is that they tend (with the help of some good old fashioned AI that nobody wants to talk about) to run multiple large language models in the background many times until they get the answers that by some criteria seem to correct.”
Our reading: AI agents tend to be more expensive because they tend to run multiple large language models in the background many times until they get answers that by some criteria seem correct.
No horizon Evidence as of
Limitation: The main evidence is one provider's description of its own system.
Assessment: SupportedSaid
Gary Marcus Observation
“Opensource models (though perhaps not Meta’s) are only a few months behind.”
Our reading: Open-source models, perhaps excluding Meta's, are only a few months behind frontier models.
No horizon Evidence as of
Limitation: One measurement organization and one index.
Assessment: SupportedSaid
Gary Marcus Observation
“And, breaking news, costs are rapidly rising too:”
Our reading: The costs of developing AI data centers, including OpenAI's Stargate, are rising rapidly.
No horizon Evidence as of
Limitation: The Stargate-specific costs Marcus pointed to were not directly checked.
Source: An entire industry is being propped up by math that is insane.
Assessment: SupportedSaid
Gary Marcus Observation
“Which is funny, because Sir Demis Hassabis [let’s call him Sir Demis Hassabis #2] , just a few months earlier in January 2026 at Davos, had something rather different to say, anticipating a somewhat slower timeline of 2031-2036:”
Our reading: Hassabis predicted AGI around 2030 in mid-2026 but gave a slower 2031-2036 timeline at Davos in January 2026.
No horizon Evidence as of
Limitation: The Davos evidence is a republication of an edited interview; Stanford evidence is reporting rather than a directly retrieved recording transcript.
Assessment: SupportedSaid
Gary Marcus Observation
“Today, most definitely not on my bingo card, President Trump signed an executive order that was very much along on these lines:”
Our reading: On 2026-06-02 President Trump signed an executive order closely resembling Marcus's proposal for preflight checks on high-impact AI models.
No horizon Evidence as of
Limitation: A reading that requires mandatory FDA-style approval would not be supported.
Assessment: SupportedSaid
Gary Marcus Observation
“Dawkins seems to imagine that since LLMs say things people do, they must be like people, and that simply does not follow.”
Our reading: It does not follow that LLMs are like people just because they say things people say.
No horizon Evidence as of
Limitation: The evidence concerns selected tasks in one model and comes from its developer.
Source: The Pope appears to understand AI better than Geoffrey Hinton does.
Assessment: SupportedSaid
Gary Marcus Observation
“The other big news per a scoop from Berber Jin at the WSJ, is that Anthropic is projecting its first (slightly) profitable quarter ever.”
Our reading: Anthropic is projecting its first, slightly profitable quarter ever, per the WSJ.
No horizon Evidence as of
Limitation: The assessment relies on a full-text secondary account because the original WSJ report was inaccessible.
Source: Checking the math behind OpenAI and Anthropic’s latest headlines
Assessment: SupportedSaid
Gary Marcus Forecast
“In my Politico “Black Swan” prediction for 2026 year, from January I made the seemingly wild prediction that,“By the end of 2026, President Trump will have begun to distance himself from the aggressively pro-AI industry policies that characterized his AI strategy in 2025””
Our reading: By the end of 2026, President Trump will have begun distancing himself from the aggressively pro-AI-industry policies of his 2025 AI strategy.
Settled before its deadline Evidence as of
Limitation: The outcome depends on treating administration implementation under a presidential order as presidential distancing.
Assessment: SupportedSaid
Gary Marcus Observation
“As I am sure you have heard by now, the backlash against AI i keep nattering on about is growing fast.”
Our reading: Public backlash against AI is growing fast as of May 2026.
No horizon Evidence as of
Limitation: The polls establish year-over-year changes in US sentiment, not a worldwide trend or a precise growth rate during May. [e1, e2]
Assessment: SupportedSaid
Gary Marcus Observation
“The US has 1,200 AI bills (roughly 150 enacted into law), but nothing that feels like a coherent AI policy.”
Our reading: The US has 1,200 AI bills, roughly 150 of them enacted into law, but nothing that feels like a coherent AI policy.
No horizon Evidence as of
Limitation: The verdict depends on reading the abbreviated post with its linked context.
Source: US AI policy is a clumsy mess. Here’s what to do about it.
Assessment: SupportedSaid
Gary Marcus Observation
“Additionally, in the new study 89.4% of agents showed drift relative to their goals after about 30 steps in their process, and 94% of agents with some form of memory-augmentation were vulnerable to poisoning attacks.”
Our reading: In the new study, 89.4% of agents drifted from goals after about 30 steps, and 94% of memory-augmented agents were vulnerable to poisoning.
No horizon Evidence as of
Limitation: The primary paper was not found. The assessment rests on news reports.
Assessment: SupportedSaid
Gary Marcus Observation
“A new study from researchers from Stanford, MIT CSAIL, Carnegie Mellon, ITU Copenhagen, and NVIDIA and Elloe AI Labs, examining 847 autonomous agent deployments drawn from healthcare, finance, customer service and code-generation, showed that 91% were vulnerable to subtle but dangerous tool-chaining attacks (see diagram below),”
Our reading: A new study found 91% of 847 autonomous agent deployments were vulnerable to tool-chaining attacks.
No horizon Evidence as of
Limitation: The primary paper was not found online. The assessment rests on news reports.
Assessment: SupportedSaid
Gary Marcus Observation
“The new paper very much amplifies what a team of AWS and Berkeley researchers showed in February, documenting similar vulnerabilities:”
Our reading: The new findings on agent vulnerabilities reinforce similar vulnerabilities documented by AWS and Berkeley researchers in February 2026.
No horizon Evidence as of
Limitation: Marcus does not name the February paper; it is identified as STAC by inference.
Assessment: SupportedSaid
Gary Marcus Observation
“It’s already splitting both parties.”
Our reading: Frustration with AI is already splitting both parties.
No horizon Evidence as of
Limitation: These examples do not establish party-wide realignment, formal fractures, or electoral consequences.
Source: The growing AI backlash
Assessment: SupportedSaid
Gary Marcus Observation
“The problem of course is that LLMs are mimics, and what they say isn’t always true.”
Our reading: LLMs are mimics and what they say is not always true.
No horizon Evidence as of
Limitation: The support for 'mimics' uses the ordinary behavioral reading specified in the contract.
Assessment: SupportedSaid
Gary Marcus Observation
“Realism re AI coding is knowing that next-word prediction gets us a surprisingly long way in writing code, but less far in making sure that code is robust.”
Our reading: Next-word prediction gets us a surprisingly long way in writing code, but less far in making sure that code is robust.
No horizon Evidence as of
Limitation: The April results come from a security vendor’s own research announcement; commercial interests warrant caution, and the packet provides no independent replication [e1].
Assessment: SupportedSaid
Gary Marcus Observation
“But that’s exactly why we still need software engineers — and why what Amodei recently said about software engineers disappearing is so absurd.”
Our reading: Software engineers are still needed, and Amodei's claim that they will disappear is absurd.
No horizon Evidence as of
Limitation: The decisive occupational count is secondary evidence.
Source: Dario Amodei, hype, AI safety, and the explosion of vibe-coded AI disasters
Assessment: SupportedSaid
Gary Marcus Observation
“AI coding tools are genuinely revolutionizing the software industry, but more than almost anything I have ever seen, they fall into the category of things you maybe shouldn’t try at home without significant prior experience:”
Our reading: AI coding tools are genuinely revolutionizing the software industry.
No horizon Evidence as of
Limitation: The reporting period precedes the post by several months.
Source: Dario Amodei, hype, AI safety, and the explosion of vibe-coded AI disasters
Assessment: SupportedSaid
Gary Marcus Observation
“That’s still true – and it very much applies in medicine.”
Our reading: Large language models are still frequently wrong yet confident, producing authoritative bullshit that shouldn't be trusted, and this applies strongly in medicine.
No horizon Evidence as of
Limitation: Support depends on the contextual evidence-period reading.
Assessment: SupportedSaid
Gary Marcus Observation
“All of this – the hallucinations, mistakes, and overconfidence – is entirely typical of LLMs, and entirely problematic in medicine.”
Our reading: Hallucinations, mistakes and overconfidence are entirely typical of LLMs and entirely problematic in medicine.
No horizon Evidence as of
Limitation: The decisive audit uses misinformation-oriented prompts and older model versions.
Assessment: SupportedSaid
Gary Marcus Observation
“Because a large fraction of the population has begun to turn to chatbots for medical advice.”
Our reading: A large fraction of the population has begun using chatbots for medical advice.
No horizon Evidence as of
Limitation: The U.S.-adult population is an interpretation, not an explicit commitment by Marcus.
Assessment: SupportedSaid
Gary Marcus Causal claim
“we should not take the apparent “confidence” of large language models — itself an illusion of how they are trained — to mean”
Our reading: The apparent confidence of LLMs is an illusion that comes from how they are trained.
No horizon Evidence as of
Limitation: The result depends on reading 'illusion' as misleading certainty rather than complete absence of information.
Assessment: SupportedSaid
Gary Marcus Observation
“Today, even after a very strong week, Oracle stands at just $170.”
Our reading: As of the post date, even after a very strong week, Oracle stock stands at just $170, compared with about $308 on September 11.
No horizon Evidence as of
Limitation: Marcus gives no quotation timestamp; the assessment uses daily closing prices.
Source: Peak absurdity, Part II
Assessment: SupportedSaid
Gary Marcus Observation
“On the darker side, it really does arm attackers to a greater degree than Mythos’s predecessors.”
Our reading: Claude Mythos Preview gives cyber attackers more capability than earlier Claude models did.
No horizon Evidence as of
Limitation: The finding concerns controlled offensive testing, not observed criminal deployment.
Source: Claude Mythos, evaluated
Assessment: SupportedSaid
Gary Marcus Observation
“Second, they show that a neurosymbolic hybrid model, which integrates both a neural networks (for pattern recognition) and a symbolic planner generalizes far better.”
Our reading: A neurosymbolic hybrid with a symbolic planner generalizes far better than VLAs on the Tower of Hanoi task.
No horizon Evidence as of
Limitation: This establishes the reported result for one simulated task and the tested configurations. It does not establish a general advantage across robotics.
Source: Even more good news for the future of neurosymbolic AI
Assessment: SupportedSaid
Gary Marcus Observation
“LLMs have persistent weaknesses around planning, reasoning and generalization.”
Our reading: LLMs have persistent weaknesses in planning, reasoning and generalization.
No horizon Evidence as of
Limitation: Generalization experiments use restricted settings.
Source: Even more good news for the future of neurosymbolic AI
Assessment: SupportedSaid
Gary Marcus Observation
“Third, they show that the neurosymbolic hybrid was vastly more efficent in terms of energy use - by nearly two orders of magnitude.”
Our reading: The neurosymbolic hybrid used nearly two orders of magnitude less energy than the VLAs in the Tufts study.
No horizon Evidence as of
Limitation: The outcome depends on interpreting the statement as a training comparison.
Source: Even more good news for the future of neurosymbolic AI
Assessment: SupportedSaid
Gary Marcus Observation
“Claude Code (in particular version 4.6), which does seem genuinely impressive to me and a number of friends I trust, still makes plenty of mistakes, and still can’t be trusted.”
Our reading: Claude Code (in particular version 4.6), which seems genuinely impressive to Marcus and a number of friends he trusts, still makes plenty of mistakes and still can't be trusted.
No horizon Evidence as of
Limitation: The controlled evaluation covers one programming language and a particular task suite.
Source: Even more good news for the future of neurosymbolic AI
Assessment: SupportedSaid
Gary Marcus Observation
“Putting things differently, Anthropic, when push came to shove, went exactly where I have said for 25 years that the field needed to go: to Neurosymbolic AI.”
Our reading: Anthropic adopted neurosymbolic AI in building Claude Code.
No horizon Evidence as of
Limitation: The result depends on Marcus's broad contextual definition of neurosymbolic AI.
Assessment: SupportedSaid
Gary Marcus Observation
“And the thing is, Claude Code is NOT a pure LLM. And it’s not pure deep learning. Not even close.”
Our reading: Claude Code is not a pure LLM and not pure deep learning.
No horizon Evidence as of
Limitation: The independent study is a preprint.
Assessment: SupportedSaid
Gary Marcus Observation
“Right now he is in trouble. People are increasingly concerned about the economics of his company,”
Our reading: As of April 2026, Sam Altman is in trouble, with people increasingly concerned about OpenAI's economics.
No horizon Evidence as of
Limitation: The result depends on reading in trouble as facing pressure and scrutiny.
Assessment: SupportedSaid
Gary Marcus Observation
“Third, there is the well-known tendency of generative AI to be sycophantic; to tell users that their ideas were the greatest ever, even when they are not.”
Our reading: Generative AI tends to be sycophantic, telling users their ideas are excellent even when they are not.
No horizon Evidence as of
Limitation: The decisive March evidence is an independent benchmark, not a peer-reviewed representative survey.
Source: In the Iran war, it looks like AI helped with operations, not strategy
Assessment: SupportedSaid
Gary Marcus Observation
“This study reinforces what Anh Totti Nguyen has been saying for a long time, in a series of underappreciated papers like Vision Language Models are Blind that I keep trying to draw attention to.”
Our reading: The new Stanford study reinforces what Anh Totti Nguyen has long been saying in a series of underappreciated papers, such as 'Vision Language Models are Blind'.
No horizon Evidence as of
Limitation: The studies test different model versions, tasks, and failure mechanisms.
Source: The mirage of visual understanding in current frontier models
Assessment: SupportedSaid
Gary Marcus Observation
“In my 2020 article, The Next Decade in AI, I foresaw all this, and urged the field to start focusing on world (cognitive) models and neurosymbolic AI.”
Our reading: In his 2020 article 'The Next Decade in AI', Marcus foresaw these developments and urged the field to start focusing on world (cognitive) models and neurosymbolic AI.
No horizon Evidence as of
Limitation: Support depends on interpreting 'foresaw all this' as a general warning, rather than a specific prediction of the 2026 Meta and xAI developments.
Source: BREAKING: Expensive new evidence that scaling is not all you need
Assessment: SupportedSaid
Gary Marcus Observation
“the fact that LLMs hallucinate (patently true)”
Our reading: LLMs hallucinate.
No horizon Evidence as of
Limitation: The evidence is a manufacturer's internal evaluation.
Assessment: SupportedSaid
Gary Marcus Observation
“It is a misapprehension of LLMs to confuse the things they say with what is true.”
Our reading: What LLMs say should not be taken as true just because they say it; treating their statements as truth misunderstands them.
No horizon Evidence as of
Limitation: The benchmark uses one person's biography and that person's manual judgments; it cannot establish a general error rate.
Assessment: SupportedSaid
Gary Marcus Observation
“Essentially the same argument applies to Claude’s constitution, which amounts to a set of built-in guardrails. OpenAI has those, too!”
Our reading: Claude's constitution is basically a set of built-in guardrails, and OpenAI's models have guardrails too.
No horizon Evidence as of
Limitation: The outcome depends on reading guardrails broadly and amounts to as shorthand.
Assessment: SupportedSaid
Gary Marcus Observation
“How about copyright? OpenAI notoriously trained on practically anything that is “publicly available” whether public domain or not.”
Our reading: OpenAI trained on practically anything publicly available, whether public domain or not.
No horizon Evidence as of
Limitation: Support applies to broad use including copyrighted material, not a numerical interpretation of “practically anything.”
Assessment: SupportedSaid
Gary Marcus Observation
“The public anger on social media is growing fast. A major boycott is afoot:”
Our reading: Public anger at OpenAI on social media is growing fast and a major boycott is underway.
No horizon Evidence as of
Limitation: No platform-wide time series establishes the precise growth rate of angry social-media activity.
Source: BREAKING: Sam Altman’s greed and dishonesty are finally catching up to him
Assessment: SupportedSaid
Gary Marcus Observation
“Using a system you shouldn’t trust with your taxes to blow people up. 2026 in a nutshell.”
Our reading: Generative AI, a system that people shouldn't trust with their taxes, is being used by certain parts of Washington to blow people up.
No horizon Evidence as of
Limitation: The result depends on reading the military phrase as operational support rather than autonomous firing.
Source: Don’t trust Generative AI to do your taxes — and don’t trust it with people’s lives
Assessment: SupportedSaid
Gary Marcus Observation
“And of course Generative AI still has problems with common sense, as an endless of stream of examples have shown, this one involving circulating on X today:”
Our reading: Generative AI still has problems with common sense.
No horizon Evidence as of
Limitation: The main systematic test comes from a commercial vendor.
Assessment: SupportedSaid
Gary Marcus Observation
“It’s one thing for the government to reject Anthropic’s terms—and entirely another to banish them permanently and, absurdly and punitively declare them a supply chain risk.”
Our reading: The US government permanently banished Anthropic and absurdly and punitively declared it a supply chain risk, going beyond merely rejecting its terms.
No horizon Evidence as of
Limitation: Formal designation followed in March.
Source: The whole thing was a scam
Assessment: SupportedSaid
Gary Marcus Observation
“The company ripped off a lot of writer’s work (per the $1.5B settlement), and recently walked back its core safety pledge.”
Our reading: Anthropic ripped off a lot of writers' work, as shown by its $1.5B settlement.
No horizon Evidence as of
Limitation: The assessment reads 'ripped off' as unauthorized acquisition, not unlawful training.
Source: The whole thing was a scam
Assessment: SupportedSaid
Gary Marcus Observation
“Anthropic refused to back down.”
Our reading: Anthropic refused to back down in its dispute with the US government over military use of its AI.
No horizon Evidence as of
Limitation: This assesses the company's position during the stated period, not the eventual outcome of negotiations.
Assessment: SupportedSaid
Gary Marcus Observation
“Although there are some practical use cases, such as coding, it is an inherently unreliable technology.”
Our reading: Generative AI has some practical uses such as coding but is an inherently unreliable technology.
No horizon Evidence as of
Limitation: Coding utility is supported here by self-reports; the experiment does not establish gains across coding tasks [e1, e3].
Assessment: SupportedSaid
Gary Marcus Causal claim
“It is leading some people into serious mental health issues.”
Our reading: Generative AI is leading some people into serious mental health issues.
No horizon Evidence as of
Limitation: One clinical case does not establish population prevalence or quantify causal risk [e1].
Assessment: SupportedSaid
Gary Marcus Observation
“They often produce workslop.”
Our reading: Large language models often produce low-quality 'workslop'.
No horizon Evidence as of
Limitation: Self-reported experience is not an independently graded output sample.
Assessment: SupportedSaid
Gary Marcus Observation
“Moreover, benchmarks are often easily gamed.”
Our reading: AI benchmarks are often easily gamed.
No horizon Evidence as of
Limitation: 'Often' has no threshold. The evidence covers prominent benchmarks, not all benchmarks.
Source: Rumors of AGI’s arrival have been greatly exaggerated
Assessment: SupportedSaid
Gary Marcus Observation
“Reports that language models have produced correct proofs for isolated open problems in mathematics, including specific Erdős problems, do not alter this assessment.”
Our reading: LLM proofs of isolated open math problems, including Erdős problems, do not show general intelligence.
No horizon Evidence as of
Limitation: The 'do not alter this assessment' part is an evaluative judgment.
Source: Rumors of AGI’s arrival have been greatly exaggerated
Assessment: SupportedSaid
Gary Marcus Observation
“Current AI systems are powerful and increasingly useful tools, but they do not exhibit the flexible, self-directed competence that the original concept of artificial general intelligence was intended to capture.”
Our reading: Current AI systems are powerful and increasingly useful tools, but they do not exhibit the flexible, self-directed competence that the original concept of AGI was intended to capture.
No horizon Evidence as of
Limitation: Most of the weight rests on one benchmark, ARC-AGI-3. Its 100% mark means matching human efficiency, not only solving the games.
Source: Rumors of AGI’s arrival have been greatly exaggerated
Assessment: SupportedSaid
Gary Marcus Observation
“Task-level performance, even when impressive, is not sufficient evidence of general intelligence.”
Our reading: Impressive task-level performance is not sufficient evidence of general intelligence.
No horizon Evidence as of
Limitation: The empirical example concerns particular systems and tasks, not every model available in February 2026.
Source: Rumors of AGI’s arrival have been greatly exaggerated
Assessment: SupportedSaid
Gary Marcus Observation
“By the standards articulated in the original definitions of artificial general intelligence—robustness across environments, reliable generalization under novelty, and autonomous goal-directed behavior—current AI systems remain limited.”
Our reading: By original AGI standards, current AI systems remain limited.
No horizon Evidence as of
Limitation: The decisive benchmark results came out about five weeks after the post.
Source: Rumors of AGI’s arrival have been greatly exaggerated
Assessment: SupportedSaid
Gary Marcus Observation
“1Some state laws are trying to address these issues, but as we know the Federal government is trying to undermine the states’ abilities to pass and enforce AI regulation.”
Our reading: The US federal government is trying to undermine states' ability to pass and enforce AI regulation.
No horizon Evidence as of
Limitation: This finding establishes an attempted restriction, not that state laws had already been invalidated or enforcement stopped.
Source: We URGENTLY need a federal law forbidding AI from impersonating humans
Assessment: SupportedSaid
Gary Marcus Observation
“Well here’s a new one, you can have OpenClaw call people, and pretend to be human.”
Our reading: OpenClaw can be used to phone people while pretending to be human, a new security risk.
No horizon Evidence as of
Limitation: Requires installation, configuration and a telephony provider.
Source: We URGENTLY need a federal law forbidding AI from impersonating humans
Assessment: SupportedSaid
Gary Marcus Observation
“Google and Anthropic have each largely caught up; various Chinese companies are closing in.”
Our reading: Google and Anthropic have largely caught up with OpenAI, and Chinese companies are closing in.
No horizon Evidence as of
Limitation: The result depends on interpreting the statement as a claim about model capabilities.
Assessment: SupportedSaid
Gary Marcus Observation
“Back on planet earth, according to December’s Remote Labor Index, only 2.5% of human “online tasks” (physical tasks were excluded) could be done by AI.”
Our reading: According to the December 2025 Remote Labor Index, AI could do only 2.5% of human online tasks.
No horizon Evidence as of
Limitation: The judgment depends on keeping the claim within the named benchmark’s scope.
Source: Promises are cheap
Assessment: SupportedSaid
Gary Marcus Observation
“The security of the autogenerated code is very much in question, as for example this article last week noted (precisely as this substack has long warned):”
Our reading: The security of AI-generated code is seriously in doubt.
No horizon Evidence as of
Limitation: One main study comes from a security vendor with a commercial interest in highlighting risk. The other is a preprint.
Source: About that Matt Shumer post that has nearly 50 million views
Assessment: SupportedSaid
Gary Marcus Observation
“LLMs write code wicked fast, but some coders are starting to report burnout, and only relatively modest gains relative to that burnout, as a new story from Connie Loizos at Tech Crunch illustrates”
Our reading: LLMs write code very fast, but some coders report burnout and only modest gains compared with that burnout.
No horizon Evidence as of
Limitation: The evidence is a small set of reports: one interviewed engineer and one anonymous forum commenter. It does not show how common burnout is among coders.
Source: About that Matt Shumer post that has nearly 50 million views
Assessment: SupportedSaid
Gary Marcus Observation
“LLMs still hallucinate and continue to make boneheaded errors.”
Our reading: As of February 2026, LLMs still hallucinate and still make boneheaded errors.
No horizon Evidence as of
Limitation: Vectara measures faithfulness in summaries only; it does not cover other kinds of hallucination.
Source: BREAKING: LLM “reasoning” continues to be deeply flawed
Assessment: SupportedSaid
Gary Marcus Observation
“And reasoning is still one of the core issues.”
Our reading: As of February 2026, reasoning is still one of the core unsolved problems for LLMs.
No horizon Evidence as of
Limitation: 'Core issue' is qualitative; no metric ranks reasoning failures against other limitations.
Source: BREAKING: LLM “reasoning” continues to be deeply flawed
Assessment: SupportedSaid
Gary Marcus Observation
“Not an all out crash, but an unmistakable sign that its ascent has run out of gas.”
Our reading: Nvidia's flat stock since August 2025 shows its rise has run out of gas, though it has not crashed.
No horizon Evidence as of
Limitation: Prices are intraday figures from Barron's, not official closes.
Source: Sam Altman and the day Nvidia’s meteoric rise came to an end