Yann LeCun's AMI Labs is selling investors on the AI problem chatbots cannot solve

Why it matters

AMI Labs turns LeCun's long-running critique of LLMs into a funded company: if world models work, the AI stack for robots, factories and engineering shifts away from chat-first systems.

Yann LeCun is using AMI Labs to make the case that the next major AI fight will be won outside the chatbot window.

In a BBC News interview, LeCun described current large language models such as ChatGPT, Claude and Gemini as useful systems for coding, math and text generation, then drew a line around what he believes they cannot become. "They're not a path towards human level or human-like intelligence, or even animal-like intelligence," he told the BBC, arguing that today's models are not built to handle real-world data.

That line is the founding thesis of Advanced Machine Intelligence Labs, the Paris-based AI research company LeCun founded after leaving Meta in 2025. AMI Labs says it is building AI systems that understand the real world, maintain persistent memory, reason and plan, and remain controllable and safe.

The BBC profile did not announce a new product. The hard financing event already happened. On March 9, TechCrunch reported that AMI Labs had raised $1.03 billion at roughly a $3.5 billion pre-money valuation, a seed round large enough to make the word "seed" feel almost procedural. The BBC, writing months later, described the raise as more than $1 billion and named Nvidia and Bezos Expeditions among the investors.

The size of the round matters because AMI Labs is asking investors to fund research before a conventional startup revenue curve appears. TechCrunch reported in March that the company starts from fundamental research and could take years to turn world models into commercial applications. It also had no plans to generate revenue for the time being, while still planning to work with prospective customers early. That is a different bargain from the applied AI startups selling copilots into existing software budgets. AMI Labs is buying time, compute and research talent for an architecture bet.

LeCun's bet is about physics, not chat

LeCun's critique has been consistent: language models learn statistical structure in text, while real intelligence needs internal models of the physical world. In the BBC interview, he used a pen balanced on its tip to explain the distinction. A child can infer that the pen will fall, while ignoring the unknowable detail of which exact direction it will go. His argument is that a useful real-world AI system should know which facts matter and which variables are noise.

AMI Labs' version of that idea centers on Joint Embedding Predictive Architecture, or JEPA, an approach LeCun proposed before starting the company. The BBC described JEPA as creating abstractions of the real world so an AI system can assess the outcomes of actions while filtering out useless information. AMI Labs says it is developing world models that learn abstract representations of sensor data and make predictions in representation space.

The company is defining a platform category before it has publicly shown a product, named paying customers or disclosed revenue.

LeCun's framing is especially pointed for robotics. The BBC quoted him saying, "LLMs are largely hopeless for robotics," and the article placed AMI Labs' work against the billions of dollars flowing into humanoid robots that still struggle with ordinary household tasks. The claim is plausible because the problem is messy in ways text generation is not. A robot stacking dishes, folding laundry or navigating a workshop has to deal with contact, occlusion, fragile objects, partial observations and consequences that cannot be corrected with a rewritten answer.

The world-model race is no longer academic

AMI Labs is entering a market where the research language has already become fundraising language. World Labs, founded by Fei-Fei Li, said in February that it had raised $1 billion in new funding from investors including AMD, Autodesk, Emerson Collective, Fidelity Management & Research Company, NVIDIA and Sea. World Labs says its first product, Marble, creates persistent 3D worlds from images, video or text.

Physical Intelligence, another company pursuing robot foundation models, open-sourced code and weights for its pi0 model in February 2025, describing it as a general-purpose robotic foundation model that can be fine-tuned for tasks such as folding laundry, cleaning a table and scooping coffee beans. The company's post framed embodiment as central to future AI systems that can understand physical interactions and reason about cause and effect.

Those companies are not interchangeable with AMI Labs. World Labs has a 3D creation product. Physical Intelligence has released robot model code. AMI Labs, by contrast, is still primarily a research wager around LeCun's architecture and team.

That difference cuts both ways. AMI Labs can claim a deeper break from LLM orthodoxy because it has not tied itself to a near-term application layer. It also has to prove that the break produces something customers can use. The March funding gives LeCun and his co-founders room to run the experiment, and it gives investors exposure to a category that could become essential if humanoid robots, autonomous industrial systems and AI-assisted engineering keep moving from demos toward deployment.

What the $1 billion buys

The money will flow to compute and talent. Building world models requires data, evaluation environments, researchers who can work below the product surface, and enough compute to test whether the architecture scales.

The open question is whether AMI Labs can turn LeCun's critique into a company before the market's patience shifts. Investors have been willing to underwrite research labs when the prize looks like a new foundation-model layer. They have also become quicker to ask where revenue will come from once the funding announcement fades. AMI Labs' disclosed investor list gives it runway and credibility. It does not answer when JEPA-based systems will outperform the LLM-centered stacks that are already being wired into robots, simulation systems and industrial workflows.

LeCun's advantage is that he is not trying to win a messaging war against chatbots on their strongest terrain. AMI Labs is making a narrower and harder claim: the AI systems that matter for the physical economy will need to understand the world well enough to plan inside it. If he is right, the most valuable AI models of the next decade may look less like fluent writers and more like machines that know when a balanced pen is about to fall.