Turing Award winner Yann LeCun says today's most celebrated AI systems are not a path to truly intelligent machines. At his Paris-based startup, Advanced Machine Intelligence Labs (AMI Labs), the former Meta chief AI scientist is building a fundamentally different kind of model designed to understand the physical world the way a rat or a toddler can.
Speaking to the BBC on the sidelines of the VivaTech conference in Paris, LeCun argued that large language models like ChatGPT, Claude, and Gemini have real uses but a hard ceiling. For more on the AI industry coverage shaping how researchers and investors think about these limits, the broader debate over what comes after the LLM era has become one of the most consequential questions in the field.
"We don't have robots that are nearly as good at understanding the physical world as a rat," LeCun said. He left Facebook-owner Meta in 2025 after a decade as its chief AI scientist and founded AMI Labs to pursue an alternative architecture.
Why LeCun Says LLMs Are Not Enough
LeCun contends that LLMs excel at well-defined, predictable tasks such as coding, solving mathematical problems, and generating text. But he argues those strengths obscure a deeper limitation: the systems manipulate statistical patterns rather than reason about reality.
"They basically just accumulate knowledge," he said of LLMs. "They can regurgitate something, you train them to regurgitate, but they're not particularly smart. They don't have an underlying understanding."
To illustrate the gap, LeCun balances a pen upright on its tip and lets go. Even a toddler knows the pen will topple, he notes, but no human would try to predict the exact direction it falls, because that outcome is genuinely uncertain. An LLM, by contrast, would attempt to generate a single prediction based on patterns in its training data. That prediction would almost certainly be wrong, because the model is not reasoning about the physical situation; it is producing what looks statistically plausible.
"LLMs are largely hopeless for robotics," LeCun said bluntly. "The claims that somehow by just scaling up LLMs, we're going to reach super human intelligence, that is simply not going to happen."
How JEPA Works
The architecture AMI Labs is developing is called Joint Embedding Predictive Architecture (JEPA). Rather than predicting the next word or pixel, JEPA creates abstractions of the real world that let it weigh the possible outcomes of an action.
Building those abstractions involves complex mathematics, but the principle is straightforward: the model filters out useless detail, leaving only useful representations of the world. In the pen example, a JEPA-based system would understand that predicting the precise direction of the fall is pointless, and would instead represent the situation at the level of abstraction that actually matters for decision-making.
The approach places AMI Labs within a growing movement around so-called world models, which aim to give AI systems an internal simulation of how environments behave, so they can plan and act rather than merely generate text.
A Competitive Field
Investors are betting that LeCun is onto something. Earlier in 2026, AMI Labs announced it had raised more than $1 billion (£760 million) in a seed funding round, with backers including US chip giant Nvidia and the fund that manages the private wealth of Amazon founder Jeff Bezos. It was one of the largest seed rounds in European startup history.
AMI Labs is far from alone in the race. At Oxford University, professor Ingmar Posner directs the Applied AI Lab and is leading a roughly ten-person team that has spent four years on an alternative approach he calls a "mechanistic world model," which structures knowledge so the AI can recall, combine, and modify it efficiently.
"My view is that the next decade will really be about systems that can explain," Posner told the BBC. "You need models that can answer questions like: What matters? What causes what? What would happen if I did something else?"
That lineage traces back to an influential 2018 paper by David Ha and Jürgen Schmidhuber, which argued that advances in compute and machine learning could let an AI learn tasks purely from a learned mental simulation of the world. The idea has since catalyzed significant research, including Google's Dreamer world model, a variant of which taught itself to collect diamonds in Minecraft by imagining future scenarios. Alphabet's DeepMind continues the work with its Genie model, London-based Wayve has a system called Gaia, and AI pioneer Fei-Fei Li founded World Labs in San Francisco in 2023 to pursue similar goals.
The Robotics Imperative
The push toward world models is being driven in large part by the robotics industry, which has poured billions of dollars into humanoid robots whose physical feats grow more impressive each year. But training those machines to safely perform household tasks such as ironing or loading a dishwasher remains difficult and expensive, and LeCun insists current AI models will never be good enough for that environment.
LeCun said AMI Labs will spend the rest of 2026 refining its model and hopes to begin deploying it in industrial settings in 2027. If that proves successful, the ambitions expand considerably.
"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.
As for what becomes of humans in that future, LeCun is sanguine. People will still decide what to build and which questions to ask, he said, while the AI works as a capable deputy.
"Our interaction with future AI systems, even if they are smarter than us, is going to be like the interaction between a captain of industry or a political leader with their staff of assistants, many of whom are smarter than they are."
Stay Ahead of AI
The shift from language models to world models could redraw the competitive map of artificial intelligence. Follow the latest AI developments as they unfold.
Read more AI news →
