Even the most advanced artificial intelligence systems available today remain, in one important respect, less capable than a common laboratory rat. That is the assessment of Yann LeCun, one of the most influential figures in modern AI research, who argues that current systems fall far short of understanding the physical world the way even simple animals do.
LeCun spent a decade at Meta, the parent company of Facebook, where he served as chief AI scientist before departing in 2025. He has since founded Advanced Machine Intelligence Labs (AMI Labs), a Paris-based venture dedicated to building a fundamentally different kind of AI system.
Speaking to BBC News on the sidelines of VivaTech, France's leading technology conference, LeCun said tools such as ChatGPT, Claude and Gemini have genuine utility but are not a route toward human-level, or even animal-level, intelligence. In his view, the architecture underpinning today's Large Language Models (LLMs) is simply not designed to process real-world data of the kind needed to navigate everyday physical tasks, such as a robot performing household chores.
The scale of investor confidence in LeCun's alternative approach is significant. AMI Labs disclosed earlier this year that it had raised more than $1 billion (£760 million) in seed funding, the earliest stage of start-up investment, from backers including US chipmaker Nvidia and the investment vehicle that manages Amazon founder Jeff Bezos's personal wealth. It ranks among the largest seed rounds ever raised by a European start-up.
LeCun's critique centres on the gap between statistical pattern-matching and genuine reasoning. LLMs excel at well-defined, predictable problems, such as writing code, solving mathematical equations or generating text, largely because they have absorbed and can reproduce patterns from vast bodies of training data. But he argues this amounts to sophisticated recall rather than understanding.
To illustrate the limitation, LeCun points to a simple physical demonstration: balance a pen upright on its tip, then let go. A toddler intuitively understands the pen will fall, but not which direction it will fall in, because there is no way to know. An LLM, by contrast, would attempt to generate a single, specific prediction based on statistical patterns drawn from its training data , a prediction almost guaranteed to be wrong, because the system has no grasp of the physical reality involved; it is only producing output that looks statistically plausible.
AMI Labs' proposed solution is a system it calls Joint Embedding Predictive Architecture, or JEPA.
Rather than attempting to predict exact outcomes, JEPA builds abstract internal representations of the world that strip away irrelevant detail, leaving the system with a working model of which outcomes are meaningful to assess. Applied to the falling-pen example, such a system would recognise that predicting the exact direction of the fall is not a productive question to answer in the first place.
The stakes for the wider robotics industry are considerable. Billions of dollars have poured into humanoid robot development, and the machines' capabilities have visibly advanced year on year. Yet training robots to reliably and safely perform everyday domestic tasks, ironing clothes or loading a dishwasher, for instance, has proven both technically difficult and expensive. LeCun is blunt about the underlying cause, telling BBC News that “LLMs are largely hopeless for robotics.”
LeCun is far from alone in this assessment. Ingmar Posner, Professor of Applied Artificial Intelligence at Oxford University and director of its Applied AI Lab, shares a similar diagnosis, though he is pursuing a different technical route. Posner, who also holds the title of Amazon Scholar, believes the next decade of AI progress will be defined by systems capable of causal reasoning, models able to identify what matters in a given situation, understand what causes what, and reason about counterfactual actions.
For roughly four years, Posner and a team of about ten researchers at Oxford have been developing an approach within a broader category of research known as World Models. Although the concept has existed in AI research for decades, the current wave of interest traces back to an influential 2018 paper by David Ha and Jurgen Schmidhuber, which argued that, given sufficient advances in machine learning and computing power, an AI system could learn to act by drawing on an internally generated, simulated understanding of its environment.
That paper helped spark a substantial body of follow-on research, including Google's Dreamer World Model. Last year, a variant of Dreamer demonstrated the approach's potential by learning to collect diamonds in the video game Minecraft, a notoriously difficult objective, purely by using imagined future scenarios to guide its decisions.
Posner describes the system his team is building as a "mechanistic world model," designed to organise knowledge so it can be recalled, combined and adapted as circumstances demand.
He is candid that predicting a development timeline is difficult, drawing a comparison to how dismissive AI researchers were, as recently as 2017 or 2018, about the prospect of a ChatGPT-like tool arriving within years rather than decades. The original version of ChatGPT launched in November 2022.
AMI Labs and Posner's Oxford lab are not working in isolation. Google's DeepMind is developing a comparable system called Genie, London-based autonomous vehicle company Wayve has built one called Gaia, and AI researcher Fei-Fei Li founded a dedicated venture, World Labs, in San Francisco in 2023 to pursue similar goals.
LeCun says AMI Labs will spend the remainder of this year refining its model, with an initial deployment in industrial settings targeted for next year. Should that prove successful, he envisions a longer-term path toward general-purpose intelligence systems capable of being applied to almost any task with minimal additional training.
Asked what role humans would play in a world where robots can operate with growing independence, LeCun suggested the essential human contribution would shift toward judgment and direction, deciding what problems are worth solving, what should be built, and what should be created. He compared the likely future relationship between people and advanced AI systems to that between a company leader or political figure and a team of specialist staff, many of whom may be more capable in their specific domains than the leader they serve.
