The most powerful AI systems on the planet cannot fold laundry. They cannot pick up a glass without being explicitly programmed to do so. They cannot look at a biopsy image and tell you which molecules are misbehaving inside the cell. They can, however, write you a poem about all of those things - which tells you something important about what language models are and, more urgently, what they are not.
That gap between language and reality is where the next phase of AI investment is heading. And according to two of the people building it - Yann LeCun, founder of AMI Labs, and Jean-Philippe Vert, CEO of Bioptimus - the architecture required to close that gap looks almost nothing like the technology that produced ChatGPT.
Speaking to Bloomberg Television on May 22, 2026, LeCun and Vert laid out the case for what researchers call world models: AI systems that understand physical and biological reality rather than just predicting the next word in a sentence. The conversation was short. The financial implications are not.
The background
To understand what world models are, it helps to understand what large language models - the technology behind ChatGPT, Claude, and Gemini - actually do. Despite what the name implies, these systems do not understand language. They are trained to predict which word is statistically likely to follow a given sequence of words, across an enormous amount of text. The result is something that reads like understanding. It is, in technical terms, a very sophisticated pattern-matching engine.
That approach works extraordinarily well for text. It works less well the moment you step outside of text.
If you ask a language model to control a robot arm, the model cannot predict what will happen next in the physical world the way it can predict the next word in a sentence. There are too many possibilities. The robot might knock the glass over. The glass might slip. The light might change. A language model asked to "predict what comes next in a video" faces essentially infinite valid answers - and its architecture has no way to handle that.
Deep learning - the broad family of techniques used to train these systems, which involves feeding vast amounts of data through layered mathematical structures called neural networks - has been the foundation of modern AI for about a decade. But deep learning applied to language and deep learning applied to the physical world require different structures at the top.
This distinction is not academic. It is the reason a string of prominent AI researchers have left the biggest technology companies on earth to build something different.
What is actually happening
LeCun left Meta in late 2025 after 12 years as its chief AI scientist. In March 2026, his new Paris-based company, AMI Labs - Advanced Machine Intelligence, raised $1.03 billion in seed funding at a pre-money valuation of $3.5 billion. That is the largest seed round in European history by a significant margin - Mistral AI, the previous holder of that title, raised 113 million euros at the same stage in 2023.
The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions. AMI has no product and no revenue timeline. The first year is entirely dedicated to fundamental research and hiring. Investors are not buying a company. They are buying a theory.
The theory is built on an architecture LeCun has been developing for nearly a decade called JEPA - Joint Embedding Predictive Architecture. Unlike a language model, which generates output one token at a time and requires enormous computational resources to predict the next word across a vocabulary, JEPA learns compressed representations of reality. It does not simulate every detail of the physical world. It builds a fast mental model of it - closer, LeCun argues, to how a brain works than how a spreadsheet works.
"I've been working on this for the better part of the last 15 years," LeCun said in the Bloomberg interview, "making fast progress over the last five."
The result, he told Bloomberg, is that "2026 is going to be the year of the world model."
The biology side of that argument is already further along. Jean-Philippe Vert co-founded Bioptimus - a Paris-based AI biotech company - after stints at Google Brain, Google DeepMind, and Owkin, a clinical AI firm. Bioptimus launched with $35 million in seed funding in early 2024 and has since released H-Optimus, a foundation model for histopathology - the study of tissue under a microscope - which is now in use by 12 of the top 20 pharmaceutical companies, according to Bioptimus.
In December 2025, Bioptimus unveiled M-Optimus, a world model for biology capable of connecting multiple modalities - linking what happens at the molecular level to what appears in tissue images to what that means for a patient's response to a drug.
"LLMs won't cure cancer," Vert said bluntly in the Bloomberg interview. "Cancer doesn't speak English."
The money trail
The capital flowing into world models is not a rounding error. According to AI2Work, global venture investment hit $297 billion in Q1 2026 alone - an all-time record - with AI capturing roughly 80 percent of every dollar invested. That is up from 55 percent of venture capital in Q1 2025.
Within that flood, world models are drawing a specific and concentrated stream. AMI Labs raised $1.03 billion in March 2026. World Labs - founded by Stanford AI researcher Fei-Fei Li - raised $500 million at a $5 billion valuation in February 2026. Google DeepMind launched Project Genie publicly, powered by Genie 3. Nvidia's Cosmos platform, trained on 20 million hours of real-world video data, surpassed 2 million downloads by early 2026.
The investment logic is not complicated. Humanoid robots - machines that can work in human environments - are the most anticipated hardware product in technology. But as LeCun noted in the Bloomberg interview, none of the companies currently building humanoid robots "has any idea how to make those robots smart enough to be useful for domestic or industrial applications that require some degree of generality and adaptability." The gap between a robot that can perform a scripted task and a robot that can handle an unexpected situation is precisely the gap that world models are designed to close.
That means investors are essentially fronting the cost of a prerequisite. Before the robot economy can exist, someone has to build the intelligence layer that makes robots adaptable. The bet is that whoever builds that layer - and builds it first - will be in the position that Nvidia currently occupies for data center AI: essential infrastructure, irreplaceable, priced accordingly.
Paris has a specific role in this story that goes beyond geography. Both AMI Labs and Bioptimus are headquartered there. Vert cited the density of talent in computer science and applied mathematics, the presence of major AI research labs, and the nature of the work itself. "When you think of deep AI, deep tech, scientific, grounded AI requiring very deep expertise," he said, Silicon Valley's advantage in fast-moving consumer products matters less than access to researchers who understand biology, mathematics, and physical systems at a fundamental level.
The European angle also reflects a funding reality. The largest seed round in European history was not raised in London or Berlin. It was raised for a company that has explicitly positioned itself against the dominant American paradigm of AI development. That is either a statement about European ambition or a data point about where American capital thinks the next architecture shift is coming from. Possibly both.
What people are doing about it
The pharmaceutical industry is not waiting. According to Drug Target Review, AI in early drug discovery has moved from a supplementary tool to a core part of target identification - the process of figuring out which biological mechanism a drug should attack. By 2026, companies are expected to run computational analysis on large biological datasets before committing to any laboratory work.
Owkin, Vert's former company, launched an agentic AI infrastructure in January 2026 built on data from more than 800 hospitals across a decade. Its Pathology Explorer agent - now integrated with Anthropic's Claude for Healthcare and Life Sciences - can identify cell types, gene expression patterns, and biomarkers directly from digital pathology slides, reducing computational time from weeks to hours.
Protein design - the process of engineering biological molecules to perform a specific function - has moved even faster. "AI models for protein design are literally used in every company today," Vert noted in the Bloomberg interview, "even though they did not exist five years ago." The reference point is AlphaFold, Google DeepMind's protein structure prediction system, which effectively solved a 50-year-old problem in structural biology in 2020 and has since been downloaded by millions of researchers.
On the robotics side, the transition is slower but the pressure is building. Jensen Huang, Nvidia's chief executive, raised the prospect at a recent industry gathering of an evolution beyond traditional data-center AI toward physical AI - a term that encompasses robots, autonomous vehicles, industrial automation, and any system that must interact with the real world rather than process text. The infrastructure being built for that transition requires a different kind of compute, different sensors, and different training data than anything currently in mass deployment.
LeCun said the first commercial applications of his methodology - through a small number of industry partners - are expected within roughly one year of AMI Labs' founding. A consumer-facing product, he estimated, is probably three to five years away.
The bottom line
The AI industry spent the last three years building systems that are extraordinarily good at language and almost useless in the physical world. Now the same investors who funded that buildout are betting that the next architecture - world models that understand physical reality rather than predict text - will require an entirely new set of tools, companies, and infrastructure. The capital is already moving. The question is whether the science can keep up with the timeline the money is buying.
Timeline
- 2020 - Google DeepMind's AlphaFold solves protein structure prediction, demonstrating AI's potential in biological sciences beyond language
- February 2024 - Bioptimus emerges from stealth with $35 million in seed funding, led by Jean-Philippe Vert, to build the first universal AI foundation model for biology
- December 2024 - Bioptimus releases H-Optimus, a histopathology foundation model adopted by 12 of the top 20 pharmaceutical companies
- December 2025 - Bioptimus unveils M-Optimus, a multimodal world model for biology connecting molecular and tissue-level data
- Late 2025 - Yann LeCun departs Meta after 12 years as chief AI scientist to found AMI Labs in Paris
- February 2026 - Fei-Fei Li's World Labs raises $500 million at a $5 billion valuation for spatial AI and world models
- January 12, 2026 - Owkin launches agentic AI infrastructure for biology at the JPM Healthcare Conference, integrating with Anthropic's Claude for Healthcare
- March 10, 2026 - AMI Labs announces $1.03 billion seed round at a $3.5 billion valuation - the largest seed round in European history
- May 22, 2026 - Yann LeCun and Jean-Philippe Vert appear on Bloomberg Television's "The Close" to discuss the transition from language AI to physical and biological AI
Summary
Who: Yann LeCun, founder of AMI Labs and former chief AI scientist at Meta, and Jean-Philippe Vert, CEO of Bioptimus and former research lead at Google Brain and Google DeepMind
What: Both researchers argue that large language models - the technology behind ChatGPT - are architecturally unsuited to physical and biological intelligence, and that a new class of AI called world models will define the next phase of the industry. AMI Labs has raised $1.03 billion to pursue this approach; Bioptimus is already deploying biology-specific world models in pharma companies.
When: The Bloomberg interview aired May 22, 2026. AMI Labs' record seed round closed in March 2026. LeCun estimates consumer-facing products are three to five years away; Vert says biology applications are already in use today.
Where: Both companies are headquartered in Paris. The investment competition includes World Labs in the United States and Google DeepMind and Nvidia globally.
Why: Language models cannot interact with the physical world in a generalizable way - they cannot make robots adaptable, cannot read biopsy images, cannot reason about molecular biology. The companies building the next architecture stand to capture the infrastructure layer of a robotics and biotech economy worth potentially trillions of dollars.