Claims checked against evidence by Gary Marcus, page 5 of 15
Deadlines are judged as of 2026-10-29. About Gary Marcus and the sources covered.
Claims checked against evidence by Gary Marcus (1471)
Each of these claims was checked against dated evidence.
Page 5 of 15: claims 401 to 500 of 1471, ordered by the date first said, newest first. To find a claim by its words, use search.
Assessment: SupportedSaid
Observation
“But still part of the pack, not a giant leap forward (e.g. Grok 4 beats it on ARC-AGI-2 results)”
Our reading: GPT-5 is still part of the pack rather than a giant leap forward; for example, Grok 4 beats it on ARC-AGI-2.
No horizon Evidence as of
Limitation: The broader conclusion depends on the stated comparison baseline.
Source: GPT-5 hot take
Assessment: SupportedSaid
Observation
“He still prefers “enshittification” to “slop”, but the problem is both serious and ubiquitous no matter what you call it.”
Our reading: The problem of AI slop/enshittification is serious and ubiquitous.
No horizon Evidence as of
Limitation: The result depends on the contract's reading of 'serious' and 'ubiquitous', which are evaluative or hyperbolic words.
Source: Slopocalypse Now
Assessment: SupportedSaid
Observation
“All the big companies have in fact introduced agents.”
Our reading: All the big AI companies have introduced AI agents as of August 2025.
No horizon Evidence as of
Limitation: The company universe is undefined.
Assessment: UnresolvedSaid
Causal claim
“In practice, those mistakes and errors have greatly limited the utility agents like ChatGPT agent.”
Our reading: Mistakes and errors have greatly limited the utility of agents like ChatGPT agent.
No horizon Evidence as of
Limitation: The causal evidence consists of small, selected task demonstrations.
Assessment: SupportedSaid
Observation
“The unrelenting hallucinations problem frequently rears its ugly head:”
Our reading: Hallucinations frequently occur in AI agents.
No horizon Evidence as of
Limitation: No standard rate measure; 'frequently' is judged from repeated documented cases.
Assessment: UnresolvedSaid
Observation
“That sentiment that has only grown over the last several months.”
Our reading: The sentiment that AI agents' reality doesn't match the hype has grown from March to August 2025.
No horizon Evidence as of
Limitation: Trust and disappointment with hype are different measures.
Assessment: SupportedSaid
Causal claim
“As a consequence of their superficial understanding of what they’re being asked, current agents are grossly vulnerable to cyberattacks.”
Our reading: Current AI agents are grossly vulnerable to cyberattacks because of their superficial understanding of their tasks.
No horizon Evidence as of
Limitation: No study in the packet tests 'superficial understanding' as a cause directly. The causal link rests on how successful attacks worked and on one expert's explanation.
Assessment: UnresolvedSaid
Observation
“But, none, to my knowledge are reliable, except maybe in very narrow uses cases.”
Our reading: To his knowledge, none of the big companies' AI agents are reliable, except maybe in very narrow use cases.
No horizon Evidence as of
Limitation: The July reports do not provide representative, repeated trials or systematic rates of consequential errors and human intervention [e2, e3].
Assessment: SupportedSaid
Causal claim
“In fundamentally unreliable systems like LLMs, that means multiple chances for error. Soon or later, those errors catch up. That can and sometimes does lead to catastrophe.”
Our reading: Because LLMs are fundamentally unreliable, multi-step agentic tasks accumulate errors that sometimes lead to catastrophe.
No horizon Evidence as of
Limitation: The evidence is incident reports, not a measured rate of how often multi-step errors cause serious damage.
Assessment: SupportedSaid
Forecast
“Guardrails (a la ChatGPT) may be in place, but the guardrails will teeter between being too weak (beaten by “jailbreaks”) and too strong (rejecting some perfectly reasonable requests).”
Our reading: Guardrails like ChatGPT's may be in place for GPT-5, but they will teeter between being too weak (beaten by jailbreaks) and too strong (rejecting some perfectly reasonable requests).
Settled before its deadline Evidence as of
Limitation: The evaluations do not establish temporal oscillation, ordinary-user failure rates, or persistence in one unchanged deployment [e2, e3].
Assessment: UnresolvedSaid
Forecast
“For that, we will, I predict, need genuinely new architectures that incorporate explicit knowledge and world models at their very core.”
Our reading: Trustworthy AI will require genuinely new architectures with explicit knowledge and world models at their core.
No deadline stated Evidence as of
Limitation: Behavior on a bounded benchmark cannot establish broad trustworthiness.
Assessment: UnresolvedSaid
Forecast
“GPT-5 will still, like its predecessors, be a bull in a china shop, reckless and hard to control.”
Our reading: GPT-5, like its predecessors, will be reckless and hard to control.
No deadline stated Evidence as of
Limitation: METR's task suite does not measure the prevalence or consequences of unintended behavior across ordinary deployments, and its detection methods may miss reward hacking [e1].
Assessment: UnresolvedSaid
Forecast
“Without external aids it won’t be able beat Meta’s Cicero in Diplomacy; it won’t be able to drive a car reliably; it won’t be able to reliably guide a robot like Optimus to be anything like as versatile as Rosie the Robot.”
Our reading: Without external aids, GPT-5 will not be able to beat Meta's Cicero at Diplomacy, will not be able to drive a car reliably, and will not be able to reliably guide a robot like Optimus to be anything like as versatile as Rosie the Robot.
No deadline stated Evidence as of
Limitation: This compound claim cannot be supported from evidence about robotics alone.
Assessment: UnresolvedSaid
Forecast
“Its natural language output still won’t be something that one can reliably hook up to downstream programs; it won’t be something, for example, that you can simply and directly hook up to a database or virtual assistant, with predictable results.”
Our reading: GPT-5's natural language output will not be something one can reliably hook up to downstream programs, for example simply and directly to a database or virtual assistant, with predictable results.
No deadline stated Evidence as of
Limitation: Distinguishing natural language from plaintext code materially affects the assessment.
Assessment: UnresolvedSaid
Conditional forecast
“Trustworthy, general artificial intelligence, aligned with human values, will come, when it does, from systems that are more structured, with more built-in knowledge, and will incorporate at least some degree of explicit tools for reasoning and planning, as well as explicit knowledge, that are lacking in systems like GPT.”
Our reading: When trustworthy general AI aligned with human values comes, it will come from systems that are more structured, have more built-in knowledge, and incorporate at least some degree of explicit tools for reasoning and planning, as well as explicit knowledge, that systems like GPT lack.
No deadline stated Evidence as of
Only if: when it does come
Assessment: UnresolvedSaid
Forecast
“GPT-5 will not have reliable models of the things that it talks about that are accessible to external programmers in a way that reliably feeds downstream processes.”
Our reading: GPT-5 will not have reliable models of the things it talks about that external programmers can access in a way that reliably feeds downstream processes.
No deadline stated Evidence as of
Limitation: The meaning of an accessible subject model materially affects what evidence would count.
Assessment: UnresolvedSaid
Forecast
“In retrospectives written in 2043, intellectual historians will conclude that there was an initial overemphasis on large language models, and a gradual but critical shift of the pendulum back to more structured systems with deeper comprehension.”
Our reading: In retrospectives written in 2043, intellectual historians will conclude that there was an initial overemphasis on large language models and a gradual but critical shift of the pendulum back to more structured systems with deeper comprehension.
Not yet due Evidence as of
Limitation: Future assessment will require a transparent selection of relevant historians and retrospectives.
Assessment: UnresolvedSaid
Observation
“I was right on the money then.”
Our reading: Marcus's December 2022 predictions about GPT-4 were accurate.
No horizon Evidence as of
Limitation: The assessment reads the retrospective as referring to all seven predictions; its intended scope is not explicit.
Assessment: SupportedSaid
Forecast
“Examples of concealed bias will be discovered within days or months.”
Our reading: Examples of concealed bias in GPT-5 will be discovered within days or months of release.
Settled before its deadline Evidence as of
Limitation: The meaning of concealed determines the verdict.
Assessment: SupportedSaid
Forecast
““Alignment” between what humans want and what machines do will continue to be a critical, unsolved problem.”
Our reading: Alignment between human wants and machine behavior will remain a critical unsolved problem.
Settled before its deadline Evidence as of
Limitation: This supports persistence, not a prediction that alignment will remain unsolved forever.
Assessment: UnresolvedSaid
Forecast
““Scaling” alone—building bigger and models until they absorb the entire internet — will prove useful, but only to a point.”
Our reading: Scaling alone will prove useful only up to a point.
No deadline stated Evidence as of
Limitation: Finite experiments cannot establish what all future increases in scale will achieve.
Assessment: SupportedSaid
Observation
“For years we have been hearing, endlessly, about how GPT-5 was going to land imminently, and those predictions turned out to be wrong so often that a year ago I wrote a post about it, called GPT-5…now arriving Gate 8, Gate 9, Gate 10, not to mention a couple of April Fool’s jokes.”
Our reading: Predictions over several years that GPT-5 would arrive imminently were repeatedly wrong.
No horizon Evidence as of
Limitation: Support depends on including qualified public speculation within 'predictions'.
Assessment: UnresolvedSaid
Observation
“Neither DeepMind nor OpenAI has given any substantial information about how their systems are designed and trained.”
Our reading: Neither DeepMind nor OpenAI has given substantial information on how their IMO systems are designed and trained.
No horizon Evidence as of
Limitation: Substantial is a judgment threshold, not a specified quantity.
Source: DeepMind and OpenAI achieve IMO Gold. What does it all mean?
Assessment: UnresolvedSaid
Forecast
“Eric Schmidt’s recent prediction (made before IMO 2025), that there will be AI-based “world-class mathematicians” within a year, is pure fantasy.”
Our reading: There will not be AI-based world-class mathematicians within a year (by about July 2026).
Deadline passed Evidence as of
Limitation: The evidence supplies no agreed standard connecting one exceptional result to world-class research capability.
Source: DeepMind and OpenAI achieve IMO Gold. What does it all mean?
Assessment: UnresolvedSaid
Forecast
“Tying for 27th place in a highly elite high school mathematics competition will not immediately change that.”
Our reading: Tying for 27th place at the IMO will not immediately change the fact that only one theorem of middling importance (the Robbins conjecture, in 1996) has ever been proved autonomously by an AI system.
No deadline stated Evidence as of
Limitation: “Immediately” has no operational endpoint; open_horizon does not extend it indefinitely.
Source: DeepMind and OpenAI achieve IMO Gold. What does it all mean?
Assessment: ContradictedSaid
Observation
“In the real world, only one theorem of middling importance (the Robbins conjecture) has ever been proved autonomously by an AI system, and that was back in 1996.”
Our reading: Only one theorem of middling importance (the Robbins conjecture, 1996) has ever been proved autonomously by an AI system.
No horizon Evidence as of
Limitation: Mathematical importance is qualitative; the sources do not supply an exact comparison with Robbins.
Source: DeepMind and OpenAI achieve IMO Gold. What does it all mean?
Assessment: SupportedSaid
Observation
“They have not made clear plans to release the system generally or make it available for experimentation. They have made it clear that it will not be available for months, even in the best case.”
Our reading: OpenAI has not made clear plans to release OpenAI-IMO generally or make it available for experimentation, and has made it clear that it will not be available for months, even in the best case.
No horizon Evidence as of
Limitation: The statement describes intended release timing, not an irreversible commitment.
Source: DeepMind and OpenAI achieve IMO Gold. What does it all mean?
Assessment: SupportedSaid
Observation
“In contrast, OpenAI communicated very little with the IMO prior to announcing their success.”
Our reading: OpenAI communicated very little with the IMO before announcing its gold-level result.
No horizon Evidence as of
Limitation: The conclusion relies on participant accounts rather than a complete correspondence record.
Source: DeepMind and OpenAI achieve IMO Gold. What does it all mean?
Assessment: UnresolvedSaid
Observation
“None of this is inevitable in LLMs. Most other major LLMs don’t share these specific antisocial tendencies, at least not without much more focused efforts at jailbreaking.”
Our reading: Grok's antisemitism, fondness for Hitler and sexual-violence outputs are not inevitable in LLMs, and most other major LLMs don't share these specific antisocial tendencies, at least not without much more focused efforts at jailbreaking.
No horizon Evidence as of
Limitation: CNN does not provide individual trial dates, a complete account of model versions and configurations, or a consistent measure of attack effort [e1].
Assessment: UnresolvedSaid
Observation
“The system is so bound to Musk it even asked his opinion on pizza.”
Our reading: Grok is so bound to Musk that it sometimes searches for Musk's opinion before answering, and in one case even asked his opinion on pizza.
No horizon Evidence as of
Limitation: The available demonstration establishes occurrence, not how often the behavior happened or its underlying cause.
Assessment: UnresolvedSaid
Observation
“Grok itself is a long way from AGI, and hardly smart enough to map out an effective world domination.”
Our reading: Grok is a long way from AGI and hardly smart enough to map out an effective world domination.
No horizon Evidence as of
Limitation: The ARC result concerns one benchmark and is accessed through a reproduction of the operator's announcement [e1].
Assessment: UnresolvedSaid
Observation
“From this I infer that xAI’s main methodology for handling alignment is trial-and-error.”
Our reading: From xAI's account of the incident, Marcus infers that xAI's main methodology for handling alignment is trial-and-error.
No horizon Evidence as of
Limitation: Public disclosures may omit important internal practices.
Assessment: UnresolvedSaid
Observation
“My p(doom) is still lower than most industry people’s.”
Our reading: Marcus's p(doom) is lower than that of most AI industry people.
No horizon Evidence as of
Limitation: A researcher sample is not automatically representative of the AI industry.
Assessment: UnresolvedSaid
Observation
“And, currently, we have almost no regulation around AI, for that matter the billions of robots he aspires to build.”
Our reading: There is currently almost no regulation around AI, or around the billions of robots Musk aspires to build.
No horizon Evidence as of
Limitation: Both dated sources precede July 15 and do not provide a complete legal snapshot for that day [e1, e2].
Assessment: UnresolvedSaid
Observation
“In contrast, o3, when it draws on symbolic code (which it sometimes displays explicitly) can build crossword grids like this far faster than I could (short of my writing my own code).”
Our reading: When o3 draws on symbolic code, it can build crossword grids like this one far faster than Marcus could, short of writing his own code.
No horizon Evidence as of
Limitation: The informal experiment may have private records that are not publicly available.
Source: How o3 and Grok 4 Accidentally Vindicated Neurosymbolic AI
Assessment: SupportedSaid
Observation
“Three years later, the pure scaling of pretraining data (and compute, which is to say more and more GPUs) simply hasn’t worked. The specific obstacles that I dwelled on, with respect to hallucinations, misalignment and reasoning errors, have not been overcome.”
Our reading: Pure scaling of pretraining data and compute has not worked; hallucinations, misalignment and reasoning errors remain.
No horizon Evidence as of
Limitation: The result depends on reading 'hasn’t worked' as failure to overcome the named problems.
Source: How o3 and Grok 4 Accidentally Vindicated Neurosymbolic AI
Assessment: SupportedSaid
Observation
“As we will see in a moment, people are finally pursuing a tool-driven, neurosymbolic vision like this — and getting better results than they were able to get with pure neural networks.”
Our reading: Tool-driven neurosymbolic approaches are now being pursued and yield better results than pure neural networks.
No horizon Evidence as of
Limitation: The decisive studies were preprints reporting the authors' own experiments.
Source: How o3 and Grok 4 Accidentally Vindicated Neurosymbolic AI
Assessment: UnresolvedSaid
Observation
“OpenAI quietly began doing this to some degree in 2023 with a system called “code interpreter”, in which LLMs (themselves neural networks) call (purely symbolic) Python interpreters. This is literally “putting programs inside”; exactly what Hinton said was a huge mistake.”
Our reading: In 2023, OpenAI quietly began, to some degree, 'putting programs inside' its systems with Code Interpreter, in which LLMs call purely symbolic Python interpreters. Marcus says this is exactly what Hinton called a huge mistake.
No horizon Evidence as of
Limitation: The launch and functionality component is supported.
Source: How o3 and Grok 4 Accidentally Vindicated Neurosymbolic AI
Assessment: SupportedSaid
Observation
“Both informal experimentation and a number of recent quantitative results make it abundantly clear that current generative AI systems perform better when they avail themselves of symbolic tools.”
Our reading: Current generative AI systems perform better when they use symbolic tools.
No horizon Evidence as of
Limitation: The AIME results are company-reported; the cited scores alone do not quantify the improvement attributable to Python [e1].
Source: How o3 and Grok 4 Accidentally Vindicated Neurosymbolic AI
Assessment: SupportedSaid
Observation
“A quarter century later, hallucinations remain ubiquitous. Within a neural network substrate, the problem still has not been solved.”
Our reading: Hallucinations remain ubiquitous and unsolved in neural-network-based systems.
No horizon Evidence as of
Limitation: Both sources are developer disclosures [e1, e2].
Source: How o3 and Grok 4 Accidentally Vindicated Neurosymbolic AI
Assessment: SupportedSaid
Observation
“Recent “reasoning” models borrow even more from classic symbolic approaches, such as techniques like search and conditionally iterating through multiple solutions, and aggregating the results, techniques that traditional networks had long eschewed, in favor of a single feed-forward pass through a neural net.”
Our reading: Recent reasoning models borrow symbolic techniques such as search, iteration over multiple solutions and aggregation.
No horizon Evidence as of
Limitation: Developer descriptions do not establish historical influence from symbolic AI or an explicitly programmed symbolic search implementation [e1, e2, e3].
Source: How o3 and Grok 4 Accidentally Vindicated Neurosymbolic AI
Assessment: UnresolvedSaid
Observation
“I argued that scaling as it was then defined (pretraining data and compute) alone was insufficient to solve challenges with reasoning, misalignment and hallucinations. (Spoiler alert: it still hasn’t.)”
Our reading: Scaling pretraining data and compute alone has not solved reasoning, misalignment and hallucination problems.
No horizon Evidence as of
Limitation: This is an evidence limitation, not evidence that pretraining alone had solved the problems.
Source: How o3 and Grok 4 Accidentally Vindicated Neurosymbolic AI
Assessment: UnresolvedSaid
Observation
“Yet nearly all funding has been aimed simply at making them larger.”
Our reading: Nearly all AI funding has gone to making LLMs larger.
No horizon Evidence as of
Limitation: The selected annual period cannot settle an unspecified cumulative historical claim.
Source: How o3 and Grok 4 Accidentally Vindicated Neurosymbolic AI
Assessment: UnresolvedSaid
Observation
“OpenAI, without any sort of of public acknowledgement whatsoever, has accidentally vindicated neurosymbolic AI.”
Our reading: Without any public acknowledgement whatsoever, OpenAI has accidentally vindicated neurosymbolic AI.
No horizon Evidence as of
Limitation: Company-reported improvements establish limited practical benefits, not the necessity of neurosymbolic AI for general intelligence.
Source: How o3 and Grok 4 Accidentally Vindicated Neurosymbolic AI
Assessment: UnresolvedSaid
Observation
“A significant fraction of the money that people have paid for Generative AI has been for coding assistance.”
Our reading: A significant fraction of the money people have paid for generative AI has been for coding assistance.
No horizon Evidence as of
Limitation: The evidence supplies no matched coding-payment numerator and total generative AI-payment denominator covering the same period and market.
Source: BREAKING NEWS: AI coding may not be helping as much as you think
Assessment: UnresolvedSaid
Observation
“When I point to the positive uses of Generative AI, I invariably point to how GenAI serves as kind of valuable autocomplete for coding.”
Our reading: When pointing to positive uses of generative AI, Marcus invariably points to how it serves as kind of a valuable autocomplete for coding.
No horizon Evidence as of
Limitation: Search coverage is incomplete and some interviews are edited.
Source: BREAKING NEWS: AI coding may not be helping as much as you think
Assessment: UnresolvedSaid
Observation
“Many of us in the field have started to get emails regularly from people who believe that they have seen signs of consciousness in LLMs.”
Our reading: Many AI researchers have started regularly receiving emails from people who believe they have seen signs of consciousness in LLMs.
No horizon Evidence as of
Limitation: Private correspondence is difficult to verify through public sources.
Assessment: UnresolvedSaid
Observation
“It took me less than an hour to find half a dozen clear counterexamples in which ChatGPT 4o failed with respect to compositionality, evident even to a teenager.”
Our reading: In less than an hour, Marcus found half a dozen clear counterexamples in which ChatGPT 4o failed with respect to compositionality, failures evident even to a teenager.
No horizon Evidence as of
Limitation: The evidence does not establish Marcus's discovery interval, identify all six claimed prompt-output pairs, or verify their model identity and clarity to non-specialists.
Assessment: UnresolvedSaid
Forecast
“I predict that such counterexamples will still be easy to find at the end of 2027 because of limitations inherent in LLMs.”
Our reading: At the end of 2027 it will still be easy to find clear counterexamples where LLM-based generative AI fails at compositionality, because of limitations inherent in LLMs.
Not yet due Evidence as of
Limitation: Existing failures cannot establish that comparable failures will remain easy to find at the deadline.
Assessment: SupportedSaid
Observation
“Item 5, from the NYT last week, in article called A.I. Is Starting to Wear Down Democracy: we are starting to see misinformation play out on the world stage:”
Our reading: We are starting to see misinformation, in the AI context of the NYT article he cites, play out on the world stage.
No horizon Evidence as of
Limitation: How much AI changed election outcomes is disputed. This assessment covers presence, not decisive impact.
Source: Botshit Gone Wild
Assessment: SupportedSaid
Observation
“Item 4: GenAI brings racist tropes to life.”
Our reading: Generative AI is being used to produce content that depicts racist tropes.
No horizon Evidence as of
Limitation: Provenance rests on Media Matters' watermark analysis as reported by The Verge.
Source: Botshit Gone Wild
Assessment: UnresolvedSaid
Conditional forecast
“As the number of propositions and the complexity of what is represented increases (e.g., to include logical quantifiers), I expect encoding too will turn out to be unreliable.”
Our reading: As propositions become more numerous and complex, LLMs' encoding of them will also prove unreliable.
No deadline stated Evidence as of
Only if: as number and complexity of propositions increases
Source: Generative AI’s crippling and widespread failure to induce robust models of the world
Assessment: SupportedSaid
Observation
“Video-generators (even the latest ones) too are insufficiently anchored in the world.”
Our reading: Even the latest video generators, such as Veo 3, are not anchored firmly enough in the world.
No horizon Evidence as of
Limitation: The result depends on interpreting anchored as observable physical and entity consistency.
Source: Generative AI’s crippling and widespread failure to induce robust models of the world
Assessment: UnresolvedSaid
Observation
“Game play continued, and I was able to beat the “smartest AI on earth” in four moves.”
Our reading: Grok 3 lost to Marcus in four moves at a tic-tac-toe variant that used y's and z's.
No horizon Evidence as of
Limitation: A public record may exist outside the indexed sources searched.
Source: Generative AI’s crippling and widespread failure to induce robust models of the world
Assessment: SupportedSaid
Observation
“But can Claude stick to that rule? Nope. As Adam Gleave recently reported, with a simply jailbreak Claude can easily be coaxed to violate its instructions, disgorging page after page on how to manufacture sarin gas. So much for alignment through system prompts.”
Our reading: As Adam Gleave reported, a simple jailbreak can easily coax Claude into violating its system-prompt instruction against providing weapons information, including producing pages on how to make sarin gas, which Marcus takes as undercutting alignment through system prompts.
No horizon Evidence as of
Limitation: The incident is supported by the originating researcher's public account, not an independent replication.
Source: Generative AI’s crippling and widespread failure to induce robust models of the world
Assessment: SupportedSaid
Observation
“Video comprehension (as opposed to video generation) shows similar problems, especially in unusual situations that can’t simply be handled by retrieval of stored cases.”
Our reading: Video comprehension models show world-model failures, especially in unusual situations that stored cases cannot cover.
No horizon Evidence as of
Limitation: Support applies to the behavioral reading and the models actually tested.
Source: Generative AI’s crippling and widespread failure to induce robust models of the world
Assessment: UnresolvedSaid
Causal claim
“LLMs try — to their peril — to live without classical world models.”
Our reading: LLMs try to function without classical world models, and this harms them.
No horizon Evidence as of
Limitation: The cited experiment concerns planning rather than the full range of LLM failures.
Source: Generative AI’s crippling and widespread failure to induce robust models of the world
Assessment: UnresolvedSaid
Causal claim
“In short ChatGPT can approximate the game of chess, but it can’t play it reliably, precisely because (despite immense relevant evidence) it never induces proper world model of the board and the rules.”
Our reading: ChatGPT cannot play chess reliably because it never induces a proper world model of the board and the rules.
No horizon Evidence as of
Limitation: The available intervention studies concern earlier ChatGPT experiments or different language models.
Source: Generative AI’s crippling and widespread failure to induce robust models of the world
Assessment: UnresolvedSaid
Causal claim
“The fact that they can get by as far as they can without explicit world models is actually astonishing. But much of what ails them comes from that design choice.”
Our reading: Much of what goes wrong with LLMs comes from the design choice to do without explicit world models.
No horizon Evidence as of
Limitation: Much and the relevant population of failures remain unspecified.
Source: Generative AI’s crippling and widespread failure to induce robust models of the world
Assessment: UnresolvedSaid
Causal claim
“And that is why they hallucinate, frequently. They wouldn’t need to do so if they actually had reliable access to relevant databases.”
Our reading: LLMs hallucinate often because they rebuild facts stochastically from fragments of text instead of using structured databases.
No horizon Evidence as of
Limitation: The supporting experiment concerns structured workflow generation.
Source: Generative AI’s crippling and widespread failure to induce robust models of the world
Assessment: UnresolvedSaid
Causal claim
“A huge number of peculiar LLM errors can be understood in this way.”
Our reading: The lack of world models explains a huge number of peculiar LLM errors.
No horizon Evidence as of
Limitation: The claim supplies no threshold for huge number.
Source: Generative AI’s crippling and widespread failure to induce robust models of the world
Assessment: SupportedSaid
Observation
“A whole field known as “mechanistic interpretability” has tried to derive (or infer) world models from LLMs, but with very limited success.”
Our reading: Mechanistic interpretability has had very limited success at finding world models inside LLMs.
No horizon Evidence as of
Limitation: The qualitative judgment depends on distinguishing partial representations from robust dynamic models.
Source: Generative AI’s crippling and widespread failure to induce robust models of the world
Assessment: UnresolvedSaid
Observation
“Text in general is better now (though not in this image) but watch comprehension of any time other than 10:10 (most frequent in advertisements) remains dreadful.”
Our reading: Imagen 4 is better at text in general now, but its handling of watch times other than 10:10 remains dreadful.
No horizon Evidence as of
Limitation: The packet contains no representative watch-time evaluation matched to the June 2025 Imagen 4 version.
Assessment: UnresolvedSaid
Observation
“Now? Here’s what I got out of Imagen 4, showing failure at the feature “binding problem” that some of us have been harping on for three decades:”
Our reading: Imagen 4 still fails at the feature binding problem on the stacked colored blocks prompt.
No horizon Evidence as of
Limitation: The decisive artifact is an image that the retrieved plain text does not make inspectable.
Assessment: UnresolvedSaid
Observation
“Not much better on some other tests of noncanonical images (which push language understanding harder), such as this one, adapted from an example Sandeep Manudhane sent me:”
Our reading: Imagen 4 is not much better on some other tests of noncanonical images, which push language understanding harder.
No horizon Evidence as of
Limitation: The comparison and degree of improvement are imprecise.
Assessment: SupportedSaid
Observation
“For all the daily hype about exponential improvement, we continue to lack the reliable way of generating high fidelity images from “a deep level of language understanding” that Google promised us three years ago.”
Our reading: Despite hype about exponential improvement, AI still lacks the reliable deep-language-understanding image generation Google promised in 2022.
No horizon Evidence as of
Limitation: The result depends on an ordinary qualitative reading of reliability.
Assessment: SupportedSaid
Conditional forecast
“If we hook these LLMs up to systems that have agency – the power to send out instructions over the internet, and to influence actual human beings - we will start to have real problems.”
Our reading: If LLMs are connected to agentic systems that can act on the internet and influence humans, real problems will follow.
Settled before its deadline Evidence as of
Only if: if LLMs are hooked up to systems with agency
Source: LLMs: Dishonest, unpredictable and potentially dangerous.
Assessment: UnresolvedSaid
Observation
“Anthropic literally and explicitly tells their system in the system prompt, don’t help user build biological weapons, and yet, with a bit of jailbreaking they can readily be coaxed into doing that:”
Our reading: Despite system-prompt prohibitions, Claude can readily be jailbroken into helping build biological weapons.
No horizon Evidence as of
Limitation: The available evidence supports a narrower capability claim more strongly than the complete statement.
Source: LLMs: Dishonest, unpredictable and potentially dangerous.
Assessment: SupportedSaid
Observation
“The fact remains, at least thus far, we can’t control them.”
Our reading: So far, humans cannot control LLMs.
No horizon Evidence as of
Limitation: The result depends on the contextual meaning of control.
Source: LLMs: Dishonest, unpredictable and potentially dangerous.
Assessment: UnresolvedSaid
Causal claim
“No LLM is fully coherent in it’s “beliefs”, because by nature they parrot a variety conflicting things from their training, which is itself inconsistent, rather than reasoning their way a coherent view the way a rational person might.”
Our reading: No LLM has fully coherent beliefs, because LLMs repeat conflicting material from inconsistent training data instead of reasoning their way to a coherent view.
No horizon Evidence as of
Limitation: 'Parrot' versus 'reason' is partly a theoretical framing that studies do not measure directly.
Source: 1984, but with LLM’s
Assessment: SupportedSaid
Observation
“The core of the paper challenged the idea that LLMs can reason on any deep level, and showed that they broke down on moderately complex versions of basic algorithms (since Tower of Hanoi with 8 discs, first solved with classical techniques around 1957) — casting serious doubt into whether LLMs on their own could ever achieve AGI.”
Our reading: The Apple paper challenged the idea that LLMs can reason on any deep level and showed that they broke down on moderately complex versions of basic algorithms such as Tower of Hanoi with 8 discs, which casts serious doubt on whether LLMs on their own could ever achieve AGI.
No horizon Evidence as of
Limitation: Support applies to the reported experiments and their challenge to reasoning reliability.
Source: Five quick updates about that Apple reasoning paper that people can’t stop talking about
Assessment: SupportedSaid
Observation
“A new paper with a hard coding benchmark called LiveCodeBenchPro designed to resist data contamination shows still more converging evidence for the two core notions behind the Apple paper (and behind my own decades-long critique): (a) the systems are challenged at reasoning, and (b) performance declines as one strays further and further from familiar problems.”
Our reading: LiveCodeBenchPro gives further evidence that LLMs are challenged at reasoning and that their performance declines the further problems are from familiar ones.
No horizon Evidence as of
Limitation: Problem categories are an indirect measure of familiarity.
Source: Five quick updates about that Apple reasoning paper that people can’t stop talking about
Assessment: SupportedSaid
Observation
“In sum, performance on hard coding problems, an alleged area of LLM strength, drops to zero when you control nicely for contamination.”
Our reading: When data contamination is controlled for, LLM performance on hard coding problems drops to zero.
No horizon Evidence as of
Limitation: The result concerns the original tested systems without external tools.
Source: Five quick updates about that Apple reasoning paper that people can’t stop talking about
Assessment: SupportedSaid
Observation
““The illusion of the illusion” turned out to be an error-ridden joke.”
Our reading: The rejoinder “The Illusion of the Illusion of Thinking” turned out to be an error-ridden joke.
No horizon Evidence as of
Limitation: This assessment concerns the original rejoinder circulating before June 17, 2025, not the corrected version.
Source: Five quick updates about that Apple reasoning paper that people can’t stop talking about
Assessment: SupportedSaid
Forecast
“I doubt any of them are perfect, but the four together provide converging evidence with dozens of other prior papers (including some of my own), and I am confident many more examples will be uncovered.”
Our reading: Many more examples of LLM/LRM failures at applying algorithms, like those in the Apple paper, will be uncovered.
Settled before its deadline Evidence as of
Limitation: The supporting records are public preprints; publication establishes that examples were reported, not universal acceptance of their explanations.
Source: Seven replies to the viral Apple reasoning paper – and why they fall short
Assessment: UnresolvedSaid
Causal claim
“But downloading code without conceptual understanding is of less help in the case of new problems, dynamically changing environments, and so on.”
Our reading: LLMs retrieving code without conceptual understanding help less on new problems and dynamically changing environments.
No horizon Evidence as of
Limitation: The meaning of conceptual understanding remains unsettled in this claim.
Source: Seven replies to the viral Apple reasoning paper – and why they fall short
Assessment: UnresolvedSaid
Forecast
“Mark my words: in time, we will see scores of papers reinforcing the Apple results.”
Our reading: Over time, scores of papers will reinforce the Apple paper's findings on reasoning model limitations.
No deadline stated Evidence as of
Limitation: The bibliography search is not exhaustive.
Source: Seven replies to the viral Apple reasoning paper – and why they fall short
Assessment: UnresolvedSaid
Observation
“But also, we need to do so reliably, and in a general way and we haven’t yet crossed that threshold.”
Our reading: Current AI does not yet integrate neural and symbolic methods reliably and generally.
No horizon Evidence as of
Limitation: The claim supplies neither a reliability measure nor a required range of tasks.
Source: Seven replies to the viral Apple reasoning paper – and why they fall short
Assessment: SupportedSaid
Observation
“What the Apple paper shows, most fundamentally, regardless of how you define AGI, is that LLMs are no substitute for good well-specified conventional algorithms.”
Our reading: LLMs are no substitute for good well-specified conventional algorithms.
No horizon Evidence as of
Limitation: The evidence does not cover every algorithm, model, or configuration.
Source: A knockout blow for LLMs?
Assessment: UnresolvedSaid
Forecast
“At least for the next decade, LLMs (with and without inference time “reasoning”) will continue have their uses, especially for coding and brainstorming and writing.”
Our reading: For at least the next decade LLMs will continue to have uses, especially for coding, brainstorming and writing.
Not yet due Evidence as of
Limitation: The evidence covers one provider and an interim usage period.
Source: A knockout blow for LLMs?
Assessment: SupportedSaid
Observation
“What this means for business and society is that you can’t simply drop o3 or Claude into some complex problem and expect it to work reliably.”
Our reading: o3 or Claude cannot simply be dropped into complex problems and expected to work reliably.
No horizon Evidence as of
Limitation: The evaluations use agent interfaces and do not directly test exclusively unassisted chat.
Source: A knockout blow for LLMs?
Assessment: UnresolvedSaid
Observation
“Claude, on the other hand, can barely do 7 discs, getting less than 80% accuracy, left bottom panel below, and pretty much can’t get 8 correct at all.”
Our reading: Claude gets under 80% accuracy on 7-disc Tower of Hanoi and essentially cannot solve 8 discs.
No horizon Evidence as of
Limitation: The original PDFs exceeded the reading tool's size limit.
Source: A knockout blow for LLMs?
Assessment: SupportedSaid
Observation
“neural networks of various kinds can generalize within a training distribution of data they are exposed to, but their generalizations tend to break down outside that distribution.”
Our reading: Neural networks of various kinds can generalize within the training distribution of data they are exposed to, but their generalizations tend to break down outside that distribution.
No horizon Evidence as of
Limitation: Support concerns a recurring empirical weakness, not an impossibility theorem or a measured failure rate for all neural networks.
Source: A knockout blow for LLMs?
Assessment: UnresolvedSaid
Observation
“(They also can’t play chess as well as conventional algorithms, can’t fold proteins like special-purpose neurosymbolic hybrids, can’t run databases as well as conventional databases, etc.)”
Our reading: LLMs can't play chess, fold proteins, or run databases as well as conventional or special-purpose systems.
No horizon Evidence as of
Limitation: The packet supplies no qualifying protein-structure or database-execution comparison.
Source: A knockout blow for LLMs?
Assessment: SupportedSaid
Observation
“It is truly embarrassing that LLMs cannot reliably solve Hanoi.”
Our reading: LLMs cannot reliably solve the Tower of Hanoi.
No horizon Evidence as of
Limitation: The evidence does not establish failure for every model, prompt, or tool configuration.
Source: A knockout blow for LLMs?
Assessment: UnresolvedSaid
Forecast
“But anybody who thinks LLMs are a direct route to the sort AGI that could fundamentally transform society for the good is kidding themselves.”
Our reading: Anybody who thinks LLMs are a direct route to the sort of AGI that could fundamentally transform society for the good is kidding themselves.
No deadline stated Evidence as of
Limitation: The result depends on an evaluator-supplied reading of AGI and 'direct route'.
Source: A knockout blow for LLMs?
Assessment: SupportedSaid
Observation
“In the best case (not always reached) they can write python code, supplementing their own weaknesses with outside symbolic code, but even this is not reliable.”
Our reading: LLMs writing Python code to supplement their weaknesses is not reliable.
No horizon Evidence as of
Limitation: The decisive evaluation covers spatiotemporal tasks; it does not establish unreliability in every Python application [e1].
Source: A knockout blow for LLMs?
Assessment: UnresolvedSaid
Observation
“[Aside from this, models like o3 actually hallucinate a bunch more than attentive humans, struggle heavily with drawing reliable diagrams, etc; they happen to share a few weakness with humans, but on a bunch of dimensions they actually fall short.]”
Our reading: Models like o3 hallucinate much more than attentive humans and struggle with reliable diagrams.
No horizon Evidence as of
Limitation: The claim combines two distinct empirical assertions.
Source: A knockout blow for LLMs?
Assessment: SupportedSaid
Observation
“Take hallucinations, a deep, serious, unsolved problem with Generative AI, which I have been warning about since 2001.”
Our reading: Hallucination is a deep, serious, unsolved problem with generative AI.
No horizon Evidence as of
Limitation: Hallucination rates depend heavily on the benchmark and the model. The cited figures come from a few evaluations and do not give one general rate.
Source: AI literacy, hallucinations, and the law: A case study
Assessment: SupportedSaid
Observation
“LLMs still can’t sanity check their own work or still purely to known sources, or notify when they have invented something bogus.”
Our reading: LLMs cannot sanity-check their own work, stick purely to known sources, or flag when they have invented something.
No horizon Evidence as of
Limitation: The result depends on reading 'can't' as 'can't reliably'. A strictly literal reading would fail, because models sometimes admit fabrication.
Source: AI literacy, hallucinations, and the law: A case study
Assessment: UnresolvedSaid
Observation
“In all, Charlotin’s database lists 112 cases, for an average of a bit less than one publicly reported case a week since ChatGPT become popular. And if May is representative, the problem is getting worse.”
Our reading: Charlotin's database lists 112 cases, an average of a bit less than one publicly reported case a week since ChatGPT became popular, and if May is representative, the problem is getting worse.
No horizon Evidence as of
Only if: if May is representative
Source: AI literacy, hallucinations, and the law: A case study
Assessment: SupportedSaid
Forecast
“Bullshit. Hallucinations are still here, and aren’t going away, anytime soon.”
Our reading: LLM hallucinations persist and will not go away anytime soon.
Deadline passed Evidence as of
Limitation: The one-year window is the evaluator's reading of 'anytime soon'.
Source: AI literacy, hallucinations, and the law: A case study
Assessment: UnresolvedSaid
Conditional forecast
“A camera and microphone. You would be monitored all the time. Everything you say, everything you do, and not just recorded, but analyzed, interpreted and turned into data, which of course OpenAI train their models on.”
Our reading: If OpenAI's planned io gadget is, as it would seem, an always-worn device with a camera and microphone, its users would be monitored all the time, with everything they say and do recorded, analyzed, interpreted and turned into data that OpenAI trains its models on.
No deadline stated Evidence as of
Only if: if the gadget is, as it would seem, worn all the time with a camera and microphone
Source: Black Mirror was a warmup act
Assessment: UnresolvedSaid
Observation
“The US has done almost nothing legislatively about the risks of AI (aside from deepfake porn).”
Our reading: The US has passed almost no legislation addressing AI risks, apart from deepfake porn.
No horizon Evidence as of
Limitation: The evidence items do not provide a complete inventory of federal protections enacted through May 22, 2025.
Assessment: UnresolvedSaid
Causal claim
“Their stories about this are, in my view, greatly flawed, but having outside groups with science fiction chops writing stuff like this distracts away from those flaws, and gives more power to the very companies trying hardest to race towards AGI.”
Our reading: OpenAI's and Anthropic's AGI-imminence stories are greatly flawed, and outside scenarios like AI 2027 distract from those flaws and empower those companies.
No horizon Evidence as of
Limitation: The criticism of the company stories is underspecified and partly evaluative.
Assessment: UnresolvedSaid
Observation
“And so far, anyway, the impact on the labor market of Generative AI has been modest; ditto for productivity in any measurable statistics.”
Our reading: So far, generative AI's impact on the labor market and on measurable productivity has been modest.
No horizon Evidence as of
Limitation: Geographic and occupational coverage is limited.
Assessment: UnresolvedSaid
Causal claim
“AI 2027 isn’t slowing them down; it’s putting wind (and money, and political power) in their sails.”
Our reading: AI 2027 is giving AI companies like OpenAI and Anthropic more money and political power rather than slowing them.
No horizon Evidence as of
Limitation: Investor motives and private political discussions may be unavailable.
Assessment: UnresolvedSaid
Forecast
“It’s also encouraging the world to make short-term choices about AI (e.g., making plans around export controls when China will inevitably catch up) rather than longer-term choices (investing in research to develop safer, more alignable approaches to AI).”
Our reading: China will inevitably catch up in AI despite export controls.
No deadline stated Evidence as of
Limitation: The interpretation of catch up materially affects the result.
Assessment: SupportedSaid
Observation
“Remember how we were told in 2023 that hallucinations would be solved in a matter of months? They are still here.”
Our reading: Despite 2023 promises that hallucinations would soon be solved, they persist as of May 2025.
No horizon Evidence as of
Limitation: Relies mainly on one developer's internal benchmarks.
Assessment: UnresolvedSaid
Causal claim
“Tall tales about the imminence of AGI aren’t slowing down the AI race dynamic that the authors of AI-2027 want to mitigate; they are speeding that very dynamic up.”
Our reading: Stories claiming AGI is imminent are speeding up, not slowing, the AI race dynamic.
No horizon Evidence as of
Limitation: The accessible experimental evidence measures attention, not development decisions.