World models represent a fundamental shift in how AI systems should learn and reason about the physical world, moving beyond the pattern-matching that powers today's large language models. A technologist who spent two decades building conversational AI systems has concluded that the current approach, while impressive at generating text, cannot deliver the kind of understanding needed for high-stakes domains like medicine, robotics, and drug discovery.

What's the Difference Between Token Prediction and World Models?

The distinction hinges on how AI systems process information. Large language models (LLMs) work by predicting the next word in a sequence, based on patterns learned from trillions of words in training data. They excel at generating fluent, coherent text. But fluency is not the same as understanding. A parrot can mimic a fire alarm without comprehending combustion.

A world model, by contrast, learns abstract representations of how environments actually work. Instead of asking "what word comes next?", it asks "what happens next in the world if I do X?" This internal simulation capability matters enormously for robotics, autonomous driving, drug discovery, and any domain where the cost of error is high and the environment is continuous rather than discrete.

The gap between these two approaches became visceral for the founder during his work in healthcare. A generative model might draft a plausible-sounding clinical note yet silently omit a contraindication because nothing in its token distribution flagged the interaction as salient. The failure mode is invisible until a pharmacist catches it, or doesn't.

Why Is a Veteran AI Founder Betting $1 Billion on This Shift?

The founder's journey reveals why this question has become urgent. In 2002, he co-founded VirtuOz in Paris, building natural language processing (NLP) engines for customer-service automation. The company landed contracts with AT&T and Symantec, was bootstrapped and relocated to San Francisco, and was acquired by Nuance in 2012. The product worked. Customers were satisfied. Yet a nagging feeling persisted: the agents could answer questions without ever forming a mental picture of what the question was about.

A year later, he started Wit.ai, a platform that let developers without deep machine learning expertise create voice interfaces. Facebook acquired Wit.ai in January 2015, and he joined Meta's AI research division, working alongside Yann LeCun and contributing to the Facebook M assistant project. These were exhilarating years with large-scale compute and world-class researchers. Yet the core architecture still relied on next-token prediction, a paradigm he was growing increasingly uncomfortable with.

In 2019, he founded Nabla, a healthcare AI startup focused on clinical documentation. By late 2024, the platform supported more than 85,000 physicians. Working in medicine taught him something no benchmark leaderboard ever could: when a system gets it wrong in a hospital, someone can be harmed. Token-level accuracy is not enough.

Late in 2025, he and Yann LeCun co-founded AMI Labs with a single mission: build AI systems that understand the real world rather than merely describing it. The seed round closed at $1.03 billion, a figure that reflects the capital intensity of training world models on high-dimensional sensory data rather than text corpora alone.

How to Evaluate Whether Your AI System Needs a World Model

• Failure Mode Assessment: If your product relies on generative text and the failure mode is "slightly wrong phrasing," an LLM is probably sufficient. If the failure mode involves physical, financial, or medical harm, start investigating causal and world-model approaches now.
• Environment Type: Token predictors excel in language-shaped tasks. World models aim at reality-shaped tasks. Determine whether your application operates in discrete, text-like domains or continuous, physical environments where prediction accuracy directly impacts safety.
• Research Tracking: Follow the research landscape. Yann LeCun's position papers on Joint Embedding Predictive Architecture (JEPA) are publicly available and worth reading regardless of technical depth. DeepMind's Genie 2 project explores generative world models for interactive environments, offering a different angle on the same underlying question.

The founder emphasizes that this is not a claim that LLMs are useless. They have clear, profitable applications. What he is claiming is that the next qualitative leap, from fluent mimicry to robust reasoning, requires a fundamentally different architecture.

What Does the Research Community Say About This Shift?

Yann LeCun, Meta's chief AI scientist, has articulated this critique in several public lectures. At the 2023 AAAI conference, he argued that autoregressive LLMs lack a "world model" capable of planning in high-dimensional, noisy environments. That talk crystallized what the founder had been feeling since the VirtuOz days.

"Large language models compress the world into low-dimensional token sequences. They produce text that reads well. They pass bar exams. They generate code that compiles. None of that, however, proves they build internal representations of causal structure," the founder explained.

Founder, AMI Labs

The shortcut of token prediction is seductive. Training on trillions of tokens yields impressive benchmarks at relatively predictable cost curves. But the moment you need a system to anticipate the physical consequences of an action, say predicting how a drug interacts with liver enzymes over 72 hours, the tokenised shortcut collapses. It hallucinates confidently. It confuses correlation with mechanism. These are not bugs to be patched; they are structural ceilings baked into the architecture itself.

What's at Stake in This Architectural Debate?

The distinction between token prediction and world models is not merely academic. It determines which problems AI can reliably solve. In robotics, an AI system that only predicts the next visual frame without understanding physics will fail catastrophically. In drug discovery, a system that generates plausible-sounding molecular structures without modeling biochemical interactions will waste time and resources. In autonomous driving, a system that pattern-matches visual inputs without simulating pedestrian behavior will cause accidents.

The founder's message to the industry is measured but clear: resist hype symmetry. The fact that LLMs are overhyped does not mean world models will deliver on every promise overnight. The honest position is measured optimism paired with engineering rigor. Whether AMI Labs succeeds or fails, the question it poses will outlast any single company: can a machine learn to simulate the world well enough to act wisely within it.