Yann LeCan’t Make a Correct Prediction
Published:
The point isn’t that smart people are sometimes wrong, or that strong opinions should be kept private. The problem is what happens when a leader in AI keeps turning the limitations of one implementation into limits of the entire field, ignores weak exponential improvements as they compound, and uses that confidence to dismiss safety work as premature. A bad prediction is harmless. Negligence about a technology moving this quickly isn’t, especially when the person making it has the influence to shape what researchers, companies, and governments take seriously. Every quote below links to the best source I could find.*
Yann LeCan't
@ylecun · Chief AI Scientist, Meta (2013–2025) · Turing Award, 2018
Yann LeCun repeatedly turns a limitation of the system in front of him into a limit on the whole field. The failure is not merely predictive. From a position of unusual influence, the same short horizons become an argument for postponing safety work while capabilities keep compounding.
What LLMs Cannot Do
Endorsing an argument that fine-tuning improved planning benchmarks by turning the task into approximate memory-based retrieval.
“Auto-Regressive LLMs can't plan (and can't really reason).”
- Yann LeCun, X (@ylecun), Sept 2023
“LLMs are useful, but they are an off ramp on the road to human-level AI. If you are a PhD student, don't work on LLMs. Try to discover methods that would lift the limitations of LLMs.”
- Yann LeCun, X (@ylecun), June 2024
“[LLMs] can do none of those or they can only do them in a very primitive way, and they don't really understand the physical world.”
- Yann LeCun, Lex Fridman Podcast #416, Mar 2024
Describing how an LLM produces an answer.
“It retrieves it because it's accumulated a lot of knowledge.”
- Yann LeCun, Lex Fridman Podcast #416, Mar 2024
Safety Can Wait
“It seems to me that before "urgently figuring out how to control AI systems much smarter than us" we need to have the beginning of a hint of a design for a system smarter than a house cat.”
- Yann LeCun, X (@ylecun), May 2024
In a longer comment supporting regulation of AI products while arguing that future systems should be made safe through careful, iterative engineering.
“AI systems are not weapons any more than a car or a computer is a weapon.”
- Yann LeCun, Facebook, Oct 2023
In the same comment, comparing advance safety work to designing a containment system for a turbojet in 1925.
“Until we have a design, it's premature to speculate about ways to make them safe.”
- Yann LeCun, Facebook, Oct 2023
“My benevolent defensive AI will be better at destroying your evil AI than your evil AI will be at hurting humans.”
- Yann LeCun, X (@ylecun), Mar 2023
Arguing that AI has no intrinsic drive to dominate, that humans set its goals, and that good AI systems could stop bad ones.
“AI systems, as smart as they might be, will be subservient to us.”
- Yann LeCun, TIME interview, Feb 2024
Economic Denial
Arguing that most jobs contain many tasks and only some will be enhanced, accelerated, or automated by AI.
“The overall cost saving is limited by the proportion of tasks that can be automated.”
- Yann LeCun, LinkedIn, Jun 2025
Gary Marcus
@GaryMarcus · Scientist, author, and longtime deep-learning critic
Gary Marcus often identifies real failures in current systems. The recurring mistake is the jump from "this model fails here" to "deep learning has reached a wall." The evidence establishes an obstacle, not a boundary.
The Scaling Wall
Published after GPT-3 and AlphaCode, eight months before ChatGPT.
“Deep Learning Is Hitting a Wall. ... We may already be running into scaling limits in deep learning.”
- Gary Marcus, Nautilus, Mar 2022
The paper tested o3-mini-high on linguistic compositionality after o1 introduced test-time compute scaling and OpenAI announced o3.
“Deep learning is hitting a wall with respect to compositionality ... a stubbornly resilient wall that cannot readily be surmounted to reach human-like compositional reasoning simply through more compute.”
- Gary Marcus, Murphy et al., arXiv:2502.10934, Feb 2025
“Game over. AGI is not imminent, and LLMs are not the royal road to getting there.”
- Gary Marcus, LinkedIn, Oct 2025
Using Tower of Hanoi failures to argue that LLMs are not a route to AGI.
“It is truly embarrassing that LLMs cannot reliably solve Hanoi. ... Anybody who thinks LLMs are a direct route ... is kidding themselves.”
- Gary Marcus, The Guardian, Jun 2025
Andrew Ng
@AndrewYNg · Founder, DeepLearning.AI · Founding lead, Google Brain
The sunny case: progress will remain gradual and alignment tools are stronger than the public realizes. Each clause is now an experiment.
Timelines
On AGI timelines and rejecting a hard-takeoff scenario.
“Maybe 50. I think it's decades out. ... Technology doesn't work like that; it advances slowly.”
- Andrew Ng, Strange Loop podcast, 2023
He immediately qualified the arithmetic while still arguing that AGI remains many decades away.
“Maybe a year ago, AGI felt 50 years away. ... another 49 years to go. These numbers are metaphorical, so don't take them too seriously.”
- Andrew Ng, Fast Company interview, Feb 2026
Safety
He compared the arguments he had heard to an unfalsifiable claim about radio waves attracting hostile aliens and called them a distraction from present harms.
“I actually spoke with quite a few people about the existential risk, and candidly, I don't get it. Many of them are very vague.”
- Andrew Ng, World Economic Forum AI Governance Summit, Nov 2023
Pointing to the difficulty of eliciting harmful instructions from ChatGPT or Bard while also supporting continued safety investment.
“The tools for aligning AI with human values, they are better than most people think. They're not perfect. ... We have better tools, more powerful, than the public probably appreciates for just telling the AI to do what we want.”
- Andrew Ng, World Economic Forum AI Governance Summit, Nov 2023
Tim Scarfe and Jeremy Budd
@MLStreetTalk · Tim Scarfe, MLST host · Jeremy Budd, University of Birmingham mathematician
Scarfe and Budd make a careful technical case, then harden properties of current training into boundaries on what the systems it produces can originate and what the labor market can replace.
Creativity and Understanding
The essay argues that current systems lack genuine creativity while explicitly leaving the door open for better architectures.
“And without this fire, AI will never venture beyond the territory it was trained on.”
- Tim Scarfe and Jeremy Budd, Why Creativity Cannot Be Interpolated, Feb 2026
Economic Denial
The argument is that producing senior engineers depends on the struggle that AI removes.
“You cannot displace the junior engineers.”
- Tim Scarfe and Jeremy Budd, Why Creativity Cannot Be Interpolated, Feb 2026
Jensen Huang
@nvidia · Founder and CEO, NVIDIA
Jensen Huang's position matters twice: he influences policy, and NVIDIA sells the infrastructure behind the AI buildout. Neither fact makes him wrong. Both raise the cost of dismissing existential and employment risks as "complete nonsense" without doing the argument.
Safety Dismissal
“I think we've done a lot of damage with very well-respected people who have painted a doomer narrative.”
- Jensen Huang, No Priors podcast, Jan 2026
“The fact that this is going to be the end of humanity, it's complete nonsense.”
- Jensen Huang, Axios, Behind the Curtain, Jul 2026
On comparisons between exporting advanced AI compute, selling nuclear weapons, and sending enriched uranium abroad.
“Comparing AI to anything that you just mentioned is lunacy. ... It's a lousy analogy. It's an illogical analogy. ... We're not enriched uranium. It's a chip, and it's a chip that they can make themselves.”
- Jensen Huang, Dwarkesh Podcast, Apr 2026
Economic Denial
“People talk about AI reducing jobs, complete nonsense.”
- Jensen Huang, NVIDIA GTC Taipei keynote, Jun 2026
* This page was assembled with help from an LLM. Due to the major limitations of LLMs, it can hallucinate. If one of these quotes is fake, the LLM has accidentally made Yann LeCun's point for him.
