LeCun Pours Cold Water: A Machine That Can Play Chess Is Not Intelligent

At Evident AI Symposium 2026, Yann LeCun once again made that rather unwelcome point: the AI field has repeatedly made the same mistake for 70 years. A machine being able to complete tasks that humans regard as intellectually complex does not mean it is smarter than a human. Computers already surpass humans on many tasks, but they are still far from human—or even animal-level—intelligence.

The clip below is from the event. In the screen, LeCun is wearing a green top; the audience is full; the setting is an Evident event.

A bit of background. LeCun is an AI pioneer and currently executive chairman of AMI Labs. The full interview is on the Evident Insights channel, 19 minutes 26 seconds, titled Yann LeCun, AMI Labs | Evident AI Symposium 2026. He discussed how large language models are still limited in reasoning and planning in the physical world; world models are seen as a possible next paradigm; he also mentioned that household robots have still not appeared.

JEPA's approach is also worth a look. According to LeCun, it makes predictions in an abstract representation space, actively ignoring unpredictable details. He used financial data as an analogy: most of it is noise, with only a little information mixed in. This judgment is sharp, because many current models are climbing leaderboards, doing better and better on measurable tasks, yet they still have not solved the problem of continually learning, planning, and acting in messy environments.

The comments section is more vivid than the main text. One commenter said that chess engines beat everyone in the 1990s, and to this day they still cannot open a door. Another brought up the TI-84 calculator: it has been superhuman for 30 years. Someone asserted that LeCun will die on this hill. Others offered another angle: AI can make more mistakes faster, so it also learns faster. Someone even @-ed Grok and asked it to prove LeCun wrong.

These replies basically lay out the controversy. On one side is task-level superhuman performance; on the other is general intelligence. The former can be piled up through narrow goals and large-scale optimization. The latter has to survive in the physical world, be able to see, move, and plan, and also learn from a small amount of information. Equating benchmark scores directly with intelligence can easily lull people into self-hypnosis. LeCun's reminder is not new, but someone has to say it again every so often, because the industry loves to translate "can do things" into "is very smart."

Full video link: https://youtu.be/DK0jypclGSk?si=aoojpJsBHzLUeDNe