Yann LeCun to academics: “You should absolutely not work on LLMs”

Yann LeCun has spent years telling the AI industry that large language models are a detour. In a lecture clip now making the rounds on X, the Turing Award winner aims the warning at a smaller and more vulnerable audience, the researchers in university labs, and drops every hedge.

“You should not work on LLM. At least if you’re in academia, you should absolutely not work on LLMs. There is nothing you can bring to the table. If you’re interested in making real progress in AI, in sort of grounded AI for the real world, if you want physical AI, don’t work on LLMs, and don’t work on generative models either.”

He said it at ETH Zurich, during his talk “World Models: Enabling the Next AI Revolution” at the Frontiers of Embodied AI lectures on 29 May. AI commentator Rohan Paul posted the clip this week as LeCun’s “latest talk”, though it is four months old. The next line got the laugh: “So, as you can probably guess, this does not make me very popular in Silicon Valley.” The slide behind him, titled “Recommendations to my fellow AI scientists”, ends in red capitals: “If you are interested in human-level AI, don’t work on LLMs.”

The first half is about compute

Strip out the provocation and the academic advice is economics. Progress on large language models is bought with training runs on clusters that only a handful of companies can pay for. A doctoral student cannot add much to a race decided by budgets, which is what “nothing you can bring to the table” means in practice. LeCun has made the point for a while: at VivaTech in Paris in 2024, as VentureBeat reported, he told students who wanted to build the next generation of AI that LLMs were in the hands of big companies.

The slide also lists what he wants the time spent on instead. Joint-embedding architectures instead of generative models. Energy-based models instead of probabilistic ones. Regularized methods instead of contrastive ones. Model-predictive control instead of reinforcement learning, which he would keep only for correcting a plan when the world does not behave as predicted. These are ideas you can test with modest hardware, and that is where a university lab still has an edge.

The second half is the real provocation

“Don’t work on generative models either” goes further than the usual case against chatbots. It rejects generation itself, the idea of building intelligence by predicting the next word or pixel, and that puts LeCun at odds with much of the physical-AI field that gets filed next to him. Last week AMD agreed to buy Fei-Fei Li’s World Labs in an all-stock deal valued at about $8.2 billion. The Register described it as a hedge against LLMs being a dead end, and noted that LeCun has said as much. Yet World Labs builds models that generate and simulate 3D environments. NVIDIA’s Cosmos 3 is a generative world model too, and Physical Intelligence builds its robot models on top of vision-language models. The money moving toward “worlds, not words” is mostly not moving toward LeCun’s version of it.

Advice from a man with a cap table

There is also the speaker’s chair to account for. LeCun left Meta and is now executive chair of AMI Labs in Paris, which raised $1.03 billion at a $3.5 billion valuation in March, Europe’s largest seed round, to build world models as an alternative to LLMs. The company is staffing up in Paris, New York, Montreal and Singapore. Every researcher who follows his advice becomes a potential hire.

That does not make the advice cynical. He held the view before AMI existed, and he paid for it in reputation long before investors paid for it in cash. But in 2026 the research recommendation and the business plan point in exactly the same direction, and students should weigh it that way.

LeCun joked that the line costs him friends in Silicon Valley. For a student choosing a thesis, the stakes are higher: a doctorate staked on the architecture he has just raised a billion dollars to prove.

Tags: Large Language Models, World models, AMI Labs, ETH Zurich, Yann LeCun