AI is 'not smart' so what's next in artificial intelligence?
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Yann LeCun left Meta in 2025 after a decade as chief AI scientist and founded Paris-based Advanced Machine Intelligence Labs (AMI Labs), which raised more than $1bn (£760m) in seed funding from Nvidia and a fund managing Jeff Bezos's private wealth — one of Europe's largest seed rounds.
- LeCun argues that LLMs like ChatGPT, Claude, and Gemini are 'not particularly smart' because they cannot reason about the physical world, illustrated by the example that an LLM would try to predict which direction a pen falls — something no human would bother guessing.
- AMI Labs is developing a system called Joint Embedding Predictive Architecture (JEPA), which builds abstractions of the real world to assess action outcomes, filtering out unpredictable detail rather than attempting a single statistical prediction.
- LeCun says current AI models are 'largely hopeless for robotics' and that the claim scaling up LLMs will reach superintelligence 'is simply not going to happen,' posing a direct challenge to the dominant AI scaling thesis.
- Ingmar Posner, professor of Applied AI at Oxford and an Amazon Scholar, leads a team of about 10 researchers developing a 'mechanistic world model' that compartmentalizes knowledge so it can be recalled, combined, and modified — a different approach within the broader World Models category.
- Other World Models efforts cited include DeepMind's Genie, London's Wayve with its Gaia system, Google's Dreamer variant (which collected diamonds in Minecraft by imagining future scenarios), and World Labs, founded by Fei-Fei Li in San Francisco in 2023.
- AMI Labs plans to refine its model through the rest of this year and hopes to deploy it in industrial settings in 2026, with LeCun envisioning eventual 'general generic intelligence systems' applicable to nearly anything with minimal training.
Why it matters: LeCun — one of AI's most prominent researchers — is publicly contradicting the industry's scaling narrative while securing $1bn in backing from Nvidia and Bezos to build an alternative. If JEPA-style architectures prove viable for robotics and physical-world reasoning, the billions flowing into LLM scaling could face a fundamental strategic challenge from a field LeCun says current models cannot enter.
