Changes, page 61 of 62

Everything we published, changed, or held back, newest first.

All changes (12,371)

Changes 12,001 to 12,200 of 12,371, newest first. Page 61 of 62.

2026-10-29
Held backBy usGary Marcus
Generative AI answers are often superficial, confabulate, fail to self-check, and regularly make basic reasoning and math errors.
The two reviewers reached different verdicts.
2026-10-29
PublishedBy usGary Marcus
Anyone claiming AGI has been achieved is either confused or hyping something.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
AI systems should solve problems by reasoning over concepts, not by surface analogy to text scraped from the internet.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
LLMs do not meet his AGI definition: they are unreliable and cannot reliably generalize to novel circumstances.
A reviewer didn’t accept one of the pieces of evidence.
2026-10-29
PublishedBy usGary Marcus
It is quite possible that no model worthy of the GPT-5 name will appear in 2025.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
AGI isn’t coming in 2025, and GPT-5 probably isn’t either.
One of the AIs read the claim differently from our summary.
2026-10-29
PublishedBy usGary Marcus
Musk's promise of 500,000 humanoid robots in three years seems very unlikely to be fulfilled.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
GPT-5 may well not be released in 2025, owing to data bottlenecks and inherent limits of LLMs.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
The media should hold those who make unrealistic AI promises to account.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
GPT-5, when it arrives, will not meet the AGI criteria Marcus has previously laid out.
It didn’t pass every check.
2026-10-29
PublishedBy usGary Marcus
o3 is a real advance beyond GPT-4 in some domains, but it is likely not the quantum leap people would expect of the name GPT-5, and at a rumored $1000 per query it is certainly not at the price point anyone might have hoped for.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
AGI will not be achieved in 2025.
The two reviewers reached different verdicts.
2026-10-29
PublishedBy usGary Marcus
Marcus's hypothesis, which he has been floating since 2022, that scaling data alone would not be enough for robust AI has become increasingly prevalent.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
OpenAI is struggling to build a model worthy of the GPT-5 name.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
As of January 2025, GPT-5 has not been released despite nearly two years of claims of exponential progress.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
The data points in Aschenbrenner's AGI projection graph are imaginary rather than real measurements.
A reviewer didn’t accept one of the pieces of evidence.
2026-10-29
PublishedBy usGary Marcus
It is imperative that we do better at reining in Silicon Valley than we have so far.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
AGI by 2027?
One of the AIs read the claim differently from our summary.
2026-10-29
PublishedBy usGary Marcus
Aschenbrenner's graph assumes all continued progress will be the same, ignoring qualitatively unsolved problems such as hallucinations, planning and reasoning, the possibility that synthetic data may not be as useful as natural data, and limits on energy distribution and compute.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
Marcus says he cannot think of a single measure on which any model existing in 2024 outpaces GPT-4 by as much as GPT-4 outpaced the 2020 state of the art, and says progress is already behind the curve.
A reviewer said our test wasn’t fair to what was actually said.
2026-10-29
Held backBy usGary Marcus
GPT-4 is not equivalent in capability to a smart high schooler.
A reviewer didn’t accept one of the pieces of evidence.
2026-10-29
PublishedBy usGary Marcus
The double Y-axis on Aschenbrenner's graph makes no sense and presupposes its own conclusion.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
As far as Marcus knows, GPT-4 on its own does not produce original work such as science-fair ideas or original computer systems.
A reviewer said our test wasn’t fair to what was actually said.
2026-10-29
PublishedBy usGary Marcus
Compared with a smart high schooler, GPT-4 can do some things better and some things worse, and it can do very few things reliably.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
We are woefully underprepared for AGI, whenever it comes, as Eliezer Yudkowsky has rightly argued for years.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
We should be worried about AGI preparedness regardless of how long until AGI arrives.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
An MIT study and six or seven later studies have shown that generative AI has not delivered real ROI for most businesses.
One AI thought this could be checked, so it needs a full check first.
2026-10-29
Held backBy usGary Marcus
So far, AI agents appear to have lots of hype but little return on investment, as with generative AI generally.
A reviewer said our test wasn’t fair to what was actually said.
2026-10-29
PublishedBy usGary Marcus
If enough other big customers are disappointed with their AI returns the way Accenture was, "the whole jig" (an unspecified current AI arrangement) will be up.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
Accenture's rollout of AI has not produced the large gains it hoped for, though AI competition may also be hurting its business.
A reviewer didn’t accept one of the pieces of evidence.
2026-10-29
Held backBy usGary Marcus
Accenture's large bet on AI does not appear to have paid off.
A reviewer said our test wasn’t fair to what was actually said.
2026-10-29
PublishedBy usGary Marcus
Claude Code will increase coders' productivity, unless technical debt cancels out the gains.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Overall, generative AI has not been delivering large gains, as many studies (MIT, McKinsey, Bain) show.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
Claude Code is a special-purpose neurosymbolic system, not a generic chatbot.
A reviewer said our test wasn’t fair to what was actually said.
2026-10-29
Held backBy usGary Marcus
Accenture's weak quarter appears to be not a blip but another example of corporate AI returns on investment falling short of expectations.
The check used the wrong kind of test for this claim.
2026-10-29
PublishedBy usGary Marcus
Tokenmaxxing (maximizing AI token use) is in decline.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
Absolutely amazing ant video augurs intellectual mayhem
The two reviewers read the claim differently.
2026-10-29
PublishedBy usGary Marcus
YouTube is likely to be ruined as a major educational resource almost overnight by AI-generated fake videos.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
More people besides the one already fooled will be fooled by fake Sora-generated videos.
The final check failed: the write-up said more than the evidence showed.
2026-10-29
PublishedBy usGary Marcus
Society faces turbulent times from AI-generated video.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
A Sora-generated ant video got the ants' anatomy wrong: it did not show the six legs real ants have.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
Soon, scammers will use Sora or open-source counterparts to make photorealistic fake videos just to sell ads.
A reviewer said our test wasn’t fair to what was actually said.
2026-10-29
PublishedBy usGary Marcus
Open-source counterparts to Sora will emerge.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Soon, a whole generation of children will be educated by AI-generated fake videos that look plausible to naive viewers but are factually incorrect.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Autonomous vehicles still are not general enough that you can just plop down a car tuned in pilot studies in Menlo Park, SF and Arizona and expect it to drive gracefully and safely in Sicily or Mumbai.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
Autonomous cars are still not reliable enough for prime time; no breakthrough has gotten them over the threshold.
The two reviewers reached different verdicts.
2026-10-29
Held backBy usGary Marcus
OpenAI's touted Rubik's cube robotics breakthrough went nowhere, and OpenAI quietly closed its robotics division a year or two later.
The two reviewers reached different verdicts.
2026-10-29
PublishedBy usGary Marcus
Marcus is pretty skeptical about OpenAI's reported Q* breakthrough.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
At least so far, OpenAI's Q* is not among the problems Marcus is worried about.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Whether Q* pans out is still an empirical question, and Marcus himself does not yet know enough about the details of Q* to judge it with certainty.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Marcus does not think Bing is already powered by OpenAI's alleged Q* breakthrough.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
It is possible that OpenAI has a breakthrough that fundamentally changes the world.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
It is not realistic that OpenAI's Q* technique has already entirely changed the world within a few months.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
OpenAI's 2019 Rubik's cube robot result, achieved on a sensor-instrumented cube in controlled lab conditions, did not guarantee generalization to the open-ended real world.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
True level 5 driverless cars are probably still many breakthroughs away.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Claimed AI breakthroughs rarely prove general enough to live up to initial expectations, often working only in some contexts.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
No current AI system comes close to reliably doing every five-hour task humans can do without error, and hallucinations and boneheaded errors remain common in everyday use.
The two reviewers reached different verdicts.
2026-10-29
PublishedBy usGary Marcus
The picture Shumer presents isn't realistic, however much people might wish it were true.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
Recent AI coding systems have changed: you can let them run more on their own than before.
The final check failed: the write-up said more than the evidence showed.
2026-10-29
PublishedBy usGary Marcus
People should not have taken Shumer's viral post about AI seriously.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
LLMs write code very fast, but some coders report burnout and only modest gains compared with that burnout.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
About that Matt Shumer post that has nearly 50 million views
One of the AIs read the claim differently from our summary.
2026-10-29
Held backBy usGary Marcus
Shumer offers no data showing that the latest coding systems can write whole complex apps without errors.
The two reviewers reached different verdicts.
2026-10-29
Held backBy usGary Marcus
When Shumer describes AIs doing longer and longer tasks on METR's task-time benchmark, he does not mention that the benchmark's criterion is 50% correct rather than 100%, or that the benchmark covers only coding and not tasks in general.
The two reviewers reached different verdicts.
2026-10-29
Held backBy usGary Marcus
LLMs are writing more code, but it is unclear whether that code is secure or trustworthy.
The final check failed: the write-up said more than the evidence showed.
2026-10-29
PublishedBy usGary Marcus
The security of AI-generated code is seriously in doubt.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Building ever larger LLMs has never succeeded in reaching AGI.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
None of what Sutskever said should actually come as a surprise, especially not to readers of Marcus's Substack or to anyone who has followed Marcus over the years.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
A trillion dollars is a terrible thing to waste
One of the AIs read the claim differently from our summary.
2026-10-29
Held backBy usGary Marcus
To a first approximation, all the big tech companies, from OpenAI, Google, Meta, xAI and Anthropic to several Chinese companies, keep doing the same experiment over and over: building ever larger LLMs in hopes of reaching AGI.
A reviewer said our test wasn’t fair to what was actually said.
2026-10-29
PublishedBy usGary Marcus
The AI investment-driven economy looks incredibly fragile.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Trillion-dollar investments in ever larger experiments aimed at AGI may be delusional.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
Larger LLMs give measurable but apparently diminishing improvements.
A reviewer said our test wasn’t fair to what was actually said.
2026-10-29
PublishedBy usGary Marcus
No LLM scaling experiment has solved the core problems of hallucination, generalization, planning and reasoning.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
If LLMs have the limits Marcus and Sutskever claim, Big Tech can reach $2T in AI revenue by 2030 only by inventing new ideas.
The final check failed: a cited source didn’t say what the write-up said.
2026-10-29
PublishedBy usGary Marcus
Most of the reasons to doubt LLMs have been known, though not widely accepted, for a very long time, and it all could have been avoided.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
The AI boom just might end up taking down the economy.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
The AI industry may have wasted a trillion dollars and several years rediscovering what cognitive science already knew.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Venture capitalists may continue to support LLM mania, at least for a while.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
If the fallout spreads wider, the cost of the LLM bet could be much more than a trillion dollars.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
In the worst case, the impact of a deflated AI bubble could be immense.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
A trillion dollars is a terrible thing to waste
One of the AIs read the claim differently from our summary.
2026-10-29
Held backBy usGary Marcus
Essentially all of what Sutskever said was in Marcus's 2018 article 'Deep learning: A Critical Appraisal' and/or his 2022 'Deep learning is hitting a wall', which argued for neurosymbolic approaches to complement neural networks, for more innate constraints, and that the Kaplan scaling laws would eventually reach diminishing returns.
The two reviewers reached different verdicts.
2026-10-29
PublishedBy usGary Marcus
Supposing for the sake of argument that LLMs have run their course and new ideas are needed, Marcus estimates, back of the envelope, the money cost of the field taking so long to recognize this at roughly a trillion dollars, much of it on Nvidia chips and massive salaries.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
A trillion dollars or more of LLM investment might be lost, with considerable economic and other collateral damage to society.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
LLMs have undermined college education, given as an example of collateral damage to society.
The two reviewers reached different verdicts.
2026-10-29
Held backBy usGary Marcus
A trillion dollars is a terrible thing to waste
One of the AIs read the claim differently from our summary.
2026-10-29
PublishedBy usGary Marcus
Research papers (Apple's reasoning paper and the chain-of-thought mirage paper) demonstrated LLM reasoning and generalization failures.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
A temporary pause on training extra large language models
One of the AIs read the claim differently from our summary.
2026-10-29
Held backBy usGary Marcus
No details of GPT-4's architecture or training set have been published.
A reviewer didn’t accept one of the pieces of evidence.
2026-10-29
PublishedBy usGary Marcus
AI development should proceed with caution until the risks and benefits are better understood.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
GPT has a truthiness problem that is not deliberate but arises because it doesn't know any better.
A reviewer didn’t accept one of the pieces of evidence.
2026-10-29
PublishedBy usGary Marcus
GPT cannot be trusted.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
GPT produces many falsehoods, but they are not really lies because GPT has no intent to deceive.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
GPT autocompletes sentences; some outputs are true because training data is vast and some are false because training data is not infinite.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
GPT's mistruths can look a lot like George Santos' lies, but that resemblance does not mean the machinery in Santos' brain has anything to do with GPT's GPUs.
The check used the wrong kind of test for this claim.
2026-10-29
PublishedBy usGary Marcus
David Sacks should hold to the position he stated in early November that AI companies should be left to their own devices without government bailouts.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
A tweet by White House AI Czar David Sacks appeared to lay the groundwork for a government bailout of overextended AI companies.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
If OpenAI fails, Google could readily take its place, so no government bailout would be required.
The two reviewers reached different verdicts.
2026-10-29
Held backBy usGary Marcus
Big problems from AI coding tools, such as AI-related outages at Amazon, have started to arrive as he predicted.
The two reviewers reached different verdicts.
2026-10-29
Held backBy usGary Marcus
“A spate of outages, including incidents tied to the use of AI coding tools”, right on schedule
The two reviewers read the claim differently.
2026-10-29
PublishedBy usGary Marcus
AI may well come to write most code.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
“A spate of outages, including incidents tied to the use of AI coding tools”, right on schedule
The final check failed: the write-up said more than the evidence showed.
2026-10-29
Held backBy usGary Marcus
Writing code, debugging it, and maintaining it a year or a decade later are different tasks, and maintaining it is even harder, as any competent coder knows.
The final check failed: the write-up said more than the evidence showed.
2026-10-29
Held backBy usGary Marcus
Some of the latest AI coding systems have done better than earlier ones, but in mission-critical systems even a small number of errors can be deadly.
The claim is hedged, so it isn’t scored.
2026-10-29
Held backBy usGary Marcus
The US now has essentially zero AI regulation.
The two reviewers reached different verdicts.
2026-10-29
Held backBy usGary Marcus
A sad day for America
One of the AIs read the claim differently from our summary.
2026-10-29
PublishedBy usGary Marcus
The US should adopt AI regulation between over- and under-regulation that fosters innovation while protecting citizens.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
A sad day for America
The two reviewers read the claim differently.
2026-10-29
PublishedBy usGary Marcus
President Trump has trampled over states' rights and left US citizens in the lurch.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
If an AI-related economic debacle occurs during Trump's term, he will be closely tied to it because of his closeness to Silicon Valley.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
People who live in the US, don't like what's happening, and don't want to leave American citizens almost completely vulnerable to the downsides of AI should contact their governor and state attorney general and let them know how they feel.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Marcus's hunch is that the AI preemption executive order will badly backfire politically on Trump, or at least on J.D. Vance.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
The new AI scaling law shell game
The final check of the write-up failed.
2026-10-29
PublishedBy usGary Marcus
Scaling alone has limited ability to advance AI.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
The new scaling laws, like the original, are empirical generalizations rather than physical laws.
A reviewer didn’t accept one of the pieces of evidence.
2026-10-29
PublishedBy usGary Marcus
The new scaling laws (post-training and inference-time) will eventually stop holding.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
People should be financially wary of the new claims about multiple AI scaling laws.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
OpenAI's own o1 research shows that scaling test-time compute helps on some problems but not others.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Marcus guesses that o1's test-time scaling works best in closed domains where a lot of reliable synthetic data can be created, such as math problems, and not so well in open-ended domains.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
The AI field must move beyond scaling and pursue genuinely new ideas.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Making LLMs, which are inherently blind to factuality, run longer rather than better is a futile way to rescue them.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
The alleged scaling law Nadella emphasized, inspired by OpenAI's o1 (the more time spent on a problem, the more accurate the result), is expensive because it by definition burns up more GPU time, and it is unreliable.
A piece of evidence came from the wrong dates.
2026-10-29
PublishedBy usGary Marcus
Scaling performance by training models longer is no longer really holding.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
In some cases o1 does not perform better than GPT-4.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
As of November 2024, a major retreat from scaling claims in the AI field has already begun.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
GPT-4 was better than GPT-3 in almost every way, whereas o1 is a far less uniform improvement.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Scaling laws aren't really laws anymore, if they ever were, and the meaning of the word 'scaling' has been greatly devalued.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
In recent weeks, much of the AI field has quietly acknowledged that recent unreleased large models are not as powerful as scaling laws predicted.
A piece of evidence came from the wrong dates.
2026-10-29
PublishedBy usGary Marcus
OpenClaw devotees should beware the robustness of AI-generated code even more than other coders, by what Marcus puts as '10x, if not 100x'.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
OpenAI President Greg Brockman sorta kinda acknowledged Marcus's point that code which compiles and passes the tests it was given is not the same as correct, secure, maintainable, well-architected software, and Marcus calls this a rare note of realism from OpenAI.
A reviewer said our test wasn’t fair to what was actually said.
2026-10-29
PublishedBy usGary Marcus
Next-word prediction gets us a surprisingly long way in writing code, but less far in making sure that code is robust.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Coders, especially inexperienced vibe coders, should be wary of the robustness of AI-generated code.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
If a billion-dollar AI system cannot be used to solve a problem (Tower of Hanoi) that Herb Simon solved with AI in 1957 and that first-semester AI students solve routinely, then the chances that models like Claude or o3 will reach AGI seem truly remote.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
LLMs may succeed on easy test sets and mislead users into thinking they have a generalizable solution when they do not.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Models like o3 hallucinate much more than attentive humans and struggle with reliable diagrams.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Marcus sarcastically suggests that getting to 'alignment' or 'safety' without reliability is a long shot.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
LLMs writing Python code to supplement their weaknesses is not reliable.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Marcus is not interested in an 'AGI' that can't do arithmetic, and he certainly would not want to entrust global infrastructure or the future of humanity to such a system.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
AGI should not be about perfectly replicating a human but should be about combining the best of both worlds: human adaptiveness with computational brute force and reliability.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Anybody who thinks LLMs are a direct route to the sort of AGI that could fundamentally transform society for the good is kidding themselves.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
LLMs cannot reliably solve the Tower of Hanoi.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Apple's new paper is pretty devastating to LLMs and is a powerful followup to a paper from many of the same authors last year.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Other forms of deep learning, especially those that work well with symbols, may eventually thrive.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
We are not going to 'extract the light cone' of the earth or 'solve physics' (Altman claims whose meaning Marcus questions) with systems that can't play Tower of Hanoi on a tower of 8 discs.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
LLMs can't play chess, fold proteins, or run databases as well as conventional or special-purpose systems.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
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.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
The LLM approach has limits that are becoming clearer every day.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Claude gets under 80% accuracy on 7-disc Tower of Hanoi and essentially cannot solve 8 discs.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
A knockout blow for LLMs?
One of the AIs read the claim differently from our summary.
2026-10-29
PublishedBy usGary Marcus
o3 or Claude cannot simply be dropped into complex problems and expected to work reliably.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
For at least the next decade LLMs will continue to have uses, especially for coding, brainstorming and writing.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
LLMs are no substitute for good well-specified conventional algorithms.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
High school students will love Galactica and will use it to fool and intimidate some of their teachers.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
A Few Words About Bullshit
One of the AIs read the claim differently from our summary.
2026-10-29
Held backBy usGary Marcus
A Few Words About Bullshit
One AI wasn’t sure who said it.
2026-10-29
PublishedBy usGary Marcus
The Galactica examples Andrew Sundstrom shared are pitch-perfect but utterly bogus imitations of science and math, presented as the real thing.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
Meta's Galactica language model frequently produces false or fabricated content.
The final check failed: a cited source didn’t say what the write-up said.
2026-10-29
PublishedBy usGary Marcus
Like Galactica, every other large language model the author has seen prevaricates (fabricates content) a lot.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
A few dark words about chatbots and death
One of the AIs read the claim differently from our summary.
2026-10-29
Held backBy usGary Marcus
OpenAI faces at least eight wrongful death lawsuits related to ChatGPT users.
A reviewer said our test wasn’t fair to what was actually said.
2026-10-29
PublishedBy usGary Marcus
By now there have been several reports of chatbot-associated deaths, both suicides and murders, enough that Wikipedia has started keeping track of them.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Chatbots apparently inducing delusions in humans is a growing problem.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Society should especially worry about chatbots influencing children, as teen deaths have shown.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
The harms Marcus warned about years ago (enshittification of the web, mass misinformation, more cyberscams) are now all happening.
A reviewer said our test wasn’t fair to what was actually said.
2026-10-29
PublishedBy usGary Marcus
Marcus is increasingly convinced that, although generative AI has its uses, on balance it is not a step forward for humanity.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
Marcus's December 2022 prediction that the first chatbot-related death would occur in 2023 proved correct.
A reviewer said our test wasn’t fair to what was actually said.
2026-10-29
PublishedBy usGary Marcus
The problem of chatbot-associated deaths is getting worse.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
By April 2023 a chatbot-related death had occurred, though causality was hard to prove.
The final check failed: the write-up said more than the evidence showed.
2026-10-29
PublishedBy usGary Marcus
Because everybody is following the same playbook of ever more massive data and servers fed into ever-larger large language models, nobody has a real technical advantage in the generative AI race.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
For all intents and purposes, the US-China generative AI race is basically a tie, with both countries serving many of their own customers using largely similar products.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Trump has come to believe that generative AI is a winner-take-all race, probably because of self-interested whisperings from Silicon Valley investors who often lean on FOMO about China to manipulate the government.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
Among American AI companies, no lead lasts more than a few months.
The final check of the write-up failed.
2026-10-29
PublishedBy usGary Marcus
Marcus does not see any plausible scenario in which either the US or China outright wins the generative AI race to the exclusion of the other.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
The US is not far ahead of China in the generative AI race.
A piece of evidence came from the wrong dates.
2026-10-29
PublishedBy usGary Marcus
Over the next few years both the US and China will build a lot of generative AI infrastructure, quite possibly significantly more than they need.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
The core misconception that is driving American AI policy
The final check failed: the write-up said more than the evidence showed.
2026-10-29
PublishedBy usGary Marcus
Nothing China can realistically do, short of full out nuclear war, is going to stop companies like Google, Microsoft, and Amazon from serving generative AI on the cloud.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
If LLMs turn out to be a dud, or are replaced by smaller, more efficient systems that don't demand such immense infrastructure, it may be all the more the case that the real winner is whichever country doesn't overextend itself to the point of financial ruin.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Generative AI models are commodities.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Building US AI policy around a fantasy that the US is somehow going to crush China in an LLM war, or that China will crush the US, is misguided.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Nothing the United States can realistically do, short of full out nuclear war, is going to stop Chinese companies from serving generative AI on China-based cloud infrastructure.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
China is currently at most a few months behind the US in generative AI.
A piece of evidence came from the wrong dates.
2026-10-29
PublishedBy usGary Marcus
The US-China lead in generative AI will be slight and flip back and forth, with neither retaining a sizeable lead for long.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Especially because chips like GPUs depreciate quickly, it may be that the real winner is whichever country doesn't overextend itself to the point of financial ruin in a foolish effort to win a race that can't be won.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
A CERN for AI and the Global Governance of AI
One of the AIs read the claim differently from our summary.
2026-10-29
PublishedBy usGary Marcus
There is no way to guarantee LLMs won't be jailbroken, and current LLMs can be jailbroken quickly and automatically.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
It is likely that at least one of the 60+ elections worldwide in 2024 will be swayed by deepfakes.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
We lack substantial controls over current AI, especially as it gets connected to real-world services and APIs.
A piece of evidence came from the wrong dates.
2026-10-29
PublishedBy usGary Marcus
Current AI cannot reliably anticipate the consequences or harms of actions it recommends.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
A CERN for AI safety and global AI safety governance should be created, provided they address a balanced range of risks.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
There is no way to guarantee that large language models reliably stick to facts.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
The world can likely benefit enormously from AI, in ways extending far beyond the productivity gains seen in recent months, though there are clearly risks too.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
AI risk discussion should focus much more on misuse of LLMs by criminals and terrorists, not just on AI takeover scenarios.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
AI need not be superintelligent to create serious problems; the real issue is control.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
We need to consider both current AI risk and AGI risk.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
People should not listen to those who deny there is any AI risk, such as Marc Andreessen, because enormous money is at stake.
Published after both AI reviews and the final check passed.
2026-10-29
PublishedBy usGary Marcus
Some AI problems may be too complex for individual labs and of insufficient financial interest to profit-driven large AI companies, which still holds true.
Published after both AI reviews and the final check passed.
2026-10-29
Held backBy usGary Marcus
A CERN for AI and the Global Governance of AI
The final check failed: a cited source didn’t say what the write-up said, and our summary didn’t match their words.
2026-10-29
PublishedBy usGary Marcus
It seems to Marcus that Anthropic's Constitutional AI idea is clever and the best of the current AI safety ideas, but still inadequate, riddled with the usual LLM problems such as unreliability, excessive dependence on similarity to training data, hallucinations and unreliable reasoning.
Published after both AI reviews and the final check passed.