Yann LeCun: Why LLMs Still Fall Short of Human-Level AI

X post by Yann LeCun (founder and president of AMI Labs, professor at NYU, former Chief AI Scientist at Meta, Turing Award), September 25, 2026, ~110 words.

By Yann LeCun — fondateur et président d'AMI Labs / September 25, 2026 / Source x.com ↗ / Reading 2 min / .md / / Machine translation

#Yann LeCun#AMI Labs#NYU#Meta#Turing Award#Noah Smith#Noahpinion#human-level AI

On September 25, 2026, Yann LeCun replies on X to economist Noah Smith. Smith had commented on a video, widely shared the same day, in which LeCun stated that simply scaling up LLMs would not reach human-level intelligence, and that the idea of a "country of geniuses in a data center" was "complete BS." To Smith, this position seemed "already falsified."

LeCun holds his ground: "Nope. It's still true." He backs his reply with four questions. Where is your domestic robot? Where is your voiture autonome de niveau 5? Where is your robot car able to learn to drive in a few hours of practice, like any 17-year-old? Where is your AI system able to understand the real world and quickly learn new skills, like a house cat?

in a few hours of practice like any 17 year old

— Yann LeCun — fondateur et président d'AMI Labs , x.com

He adds a concession. There is "no doubt" that AI will eventually become as intelligent as humans in every domain. But he sets two caveats. First, "we're still far from that," even though AI and computing surpass humans on an ever-increasing number of tasks. Second, that intelligence will not rest on LLMs, even though they will have a role to play, for instance as a text interface.

In a second post ("Also:"), LeCun points back to a reply from September 20 addressed in part to Geoffrey Hinton: he recalled having said that "autoregressive LLMs, in and of themselves, will not lead to human-level AI," and that this statement remains "completely true."

The post contains no figures, no timeline, and no description of the alternative architecture. Its argument rests on examples of learning efficiency: a human or an animal picks up a new skill with little experience and transfers it to unfamiliar situations, which current systems do not do. The driving example sums up the gap: a few hours of practice are enough for a teenager, while unrestricted self-driving still has not been achieved.

The text separates two claims often conflated in public debate: AI surpasses humans on an increasing number of tasks, and AI remains far from general human intelligence. LeCun holds both at once.

Key takeaways

The balancing sentence to remember. AI « surpasses humans in an ever-increasing number of tasks » and « we're still far from » human intelligence. Both hold at once. It's a useful point to make in a talk against the "AI will surpass humans" narrative: yes on tasks, no on general intelligence.

The driving example. a 17-year-old learns to drive in a few hours of practice (in France, the B license requires a minimum of 20 hours of supervised driving). Years of development and millions of kilometers of data have not produced a voiture autonome de niveau 5 that could drive anywhere without restriction, including in scenes never seen before. The gap is about learning efficiency, not raw performance.

The house cat. as a benchmark: understanding the physical world and quickly learning a new skill, which no current system does, according to LeCun. He is aiming at out-of-distribution generalization, not linguistic competence.

The nuance on LLMs. LeCun does not say they are useless, he says they will not be enough « in and of themselves » and that they will keep a text-interface role. The word "autoregressive" matters in his September 20 formulation.

What the post does not contain. no timeline, no measurement, no description of the alternative architecture (the world models he advocates elsewhere, via AMI Labs, are not named here). To be cited as the position of a leading researcher, not as a demonstration.

Context of the debate. the video that triggered the exchange ("no way in hell" of reaching AGI by scaling up LLMs; "country of geniuses in a data center" called « complete BS ») was circulating the same day; Noah Smith considered it already falsified. For the September 2026 sequence on the replacement narrative, see [[mhalla-what-if-the-future-does-not-need-you-2026-09-20]].

Attributed claims

we are still far from human-level AI

— Yann LeCun

AI and computing surpass humans in a growing number of tasks

— Yann LeCun

autoregressive LLMs, by themselves, will not lead to human-level AI

— Yann LeCun

no car can learn to drive in a few hours of practice like a 17-year-old teenager

— Yann LeCun

no AI system understands the real world or quickly learns new skills like a domestic cat

— Yann LeCun

Yann LeCun: Why LLMs Still Fall Short of Human-Level AI

Summary

Key takeaways

Attributed claims