Turing Award winner Yann LeCun told university researchers at ETH Zurich that they should avoid working on large language models because they have nothing to contribute to a field dominated by massive corporate budgets. “Dere bør ikke jobbe med LLM-er,” LeCun stated during his address titled “World Models: Enabling the Next AI Revolution” on May 29, according to coverage of the event.
Financial Realities of Modern Dataclusters
The core of the warning centers on compute economics rather than pure provocation. Progress in large language models relies on training runs executed on massive dataclusters that only a small number of technology corporations can afford to finance. A doctoral student cannot make a meaningful impact in a race governed strictly by enterprise budgets. This dynamic explains the Turing laureate’s blunt assessment that academic labs lack the resources to compete in traditional generative AI training.
Previous remarks from the speaker echo this assessment. At VivaTech in Paris in 2024, he advised students aspiring to build the next generation of artificial intelligence that large language models belong firmly in the hands of major corporations. University laboratories are instead advised to direct their limited hardware toward architectures featuring joint embedding rather than generative models, and energy-based models rather than probabilistic ones.
Competing Visions for Embodied Intelligence
The rejection of generative models extends beyond chatbot technology to the foundational idea of building intelligence simply by predicting the next word or pixel. This stance places the researcher at odds with significant portions of the physical AI sector. AMD agreed to acquire Fei-Fei Li’s World Labs in a stock transaction valued around 8.2 milliarder dollar, a move described by industry reporting as a hedge against large language models proving to be a dead end.
Despite that shared skepticism toward traditional language architecture, World Labs continues to develop systems that generate and simulate three-dimensional environments. NVIDIA maintains a similar path with Cosmos 3 as a generative world model, while Physical Intelligence constructs robot architectures on top of combined vision and language systems. Substantial capital flows toward companies building virtual worlds and simulated environments, though much of that funding does not support the specific architectural approach advocated by the former Meta executive.
Venture Backing for Alternative Architectures
The academic guidance aligns closely with the commercial strategy of AMI Labs in Paris, where the speaker serves as working chairman. In March, the company hentet inn 1.03 milliarder dollar til en verdsettelse på 3,5 milliarder dollar in what marked Europe’s largest seed round. The funding supports the development of world models designed as an alternative to large language models, with staff expanding across offices in Paris, New York, Montreal, and Singapore.
Students selecting a research focus now face a stark choice between conventional transformer architectures and the alternative frameworks championed by venture-backed startups. For an academic researcher drafting a dissertation topic, the stakes involve aligning career output with methodologies that attract billions in private investment or remaining tied to hardware-constrained paths.
Yann LeCun Delivers Warnings at ETH Zurich
When and where did Yann LeCun deliver his warnings against large language models?
The remarks were presented at ETH Zurich on May 29 during a lecture titled “World Models: Enabling the Next AI Revolution” as part of the Frontiers of Embodied AI lecture series.
What funding amount did AMI Labs raise for world model development?
AMI Labs secured hentet inn 1.03 milliarder dollar in March at a verdsettelse på 3,5 milliarder dollar, representing Europe’s largest seed round.
Which alternative modeling techniques does the researcher recommend for academic labs?
The presentation recommended architectures using joint embedding over generative models, energy-based models over probabilistic methods, regularized approaches over contrastive ones, and model-predictive control over reinforcement learning.