Yann LeCun, a leading figure in artificial intelligence, left his role as chief AI scientist at Meta in 2025 and founded Advanced Machine Intelligence Labs (AMI Labs). The Paris-based startup has already raised more than $1bn from investors including Nvidia and the fund managing Jeff Bezos's private wealth, marking one of Europe's largest seed funding rounds. AMI Labs is building a new type of AI that moves beyond large language models (LLMs) like ChatGPT, which LeCun says will never handle real-world complexity.
The limits of large language models
LLMs excel at coding, math problems and text generation because those tasks are well defined and predictable. LeCun argues that these systems simply regurgitate statistically plausible responses without an underlying understanding of physical reality. He demonstrated the gap by balancing a pen on its tip: a toddler knows it will topple, but no human would guess the fall direction. An LLM, trained on patterns, would likely produce a single wrong prediction because it does not reason about physics.
"They're not a path towards human level or human-like intelligence, or even animal-like intelligence, because they cannot deal with real world data, they just are not built for that," LeCun said at VivaTech, France's leading technology conference. He added that LLMs are "largely hopeless for robotics."
A new architecture for real-world AI
AMI Labs is developing a system called Joint Embedding Predictive Architecture (JEPA). Instead of generating exact predictions from statistical patterns, JEPA creates abstractions of the real world that filter out useless information. In the pen example, the system would understand that predicting the fall direction is pointless. That flexibility is critical for robotics, where billions of dollars are being invested in humanoid robots but training them for household tasks remains difficult and expensive.
LeCun said AMI Labs will spend the rest of 2025 refining the model and aims to deploy it in industrial settings next year. "Eventually down the line we'll have sort of general generic intelligence systems that can be applied to just about anything in the world with minimal training or fine tuning," he said. Despite the autonomous potential, he believes humans will still define what questions to ask and what to build, comparing the relationship to that between a leader and a staff of smarter assistants.
World models gain traction
LeCun's approach fits a broader research category called world models, which teach AI to learn through internal simulations of environments rather than static pattern matching. Ingmar Posner, professor of Applied Artificial Intelligence at Oxford University, leads a team building what he calls a "mechanistic world model" that structures knowledge so an AI can recall, combine and modify it efficiently. Posner said the next decade will focus on systems that can explain what matters, what causes what, and what would happen under different actions.
The idea received a boost from a 2018 paper by David Ha and Jurgen Schmidhuber, which showed that advances in compute power allow an AI to learn purely from a mental simulation of the world. Google's Dreamer model later used this concept to collect diamonds in Minecraft by imagining future scenarios. Other efforts include DeepMind's Genie, Wayve's Gaia, and World Labs, founded by AI pioneer Fei-Fei Li in 2023. Posner cautions that timelines are unpredictable, noting that before 2022 many researchers thought a ChatGPT-like system was decades away.
Why this matters for IT and development
For IT and development teams tracking the next wave of AI, the pivot away from LLM scaling toward architectures like JEPA will shift how models are trained, validated, and integrated into production systems. AI Agents & Automation in physical environments require models that handle ambiguity and real-world physics, not just text generation. As investment floods into world models and robotics, professionals with skills in AI for IT & Development will need to evaluate when LLMs are appropriate and when only a more flexible architecture can deliver safe, reliable performance outside of narrowly defined tasks.
Your membership also unlocks:
