I spent roughly two decades building systems that parse human language. Virtual agents, voice interfaces, clinical assistants — each project pushed me a little closer to a question I kept dodging: do these machines actually understand anything, or are they just very good at pattern-matching over tokenised sequences? This article is the tutorial I wish I had read ten years ago, back when I still believed statistical fluency was a reliable proxy for comprehension.
The Observation That Started Everything
In 2002, I co-founded VirtuOz in Paris. We built proprietary NLP engines for customer-service automation and eventually landed contracts with AT&T and Symantec. We bootstrapped the company, relocated to San Francisco, and in 2012 Nuance acquired us. The product worked. Customers were satisfied. But a nagging feeling persisted: our agents could answer questions without ever forming a mental picture of what the question was about.
A year later I started Wit.ai, a platform that let developers without deep ML expertise create voice interfaces. Facebook acquired Wit.ai in January 2015. I joined Meta’s AI research division, worked alongside Yann LeCun, and contributed to the Facebook M assistant project. These were exhilarating years — large-scale compute, world-class researchers, ambitious scope. Yet the core architecture still relied on next-token prediction, a paradigm I was growing increasingly uncomfortable with.
Why Token Prediction Feels Like a Dead End
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. A parrot can mimic a fire alarm without understanding combustion.
The shortcut 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.
Yann LeCun has articulated this critique in several public lectures, notably at the 2023 AAAI conference, where he argued that autoregressive LLMs lack a “world model” capable of planning in high-dimensional, noisy environments. I was in the audience. That talk crystallised what I had been feeling since the VirtuOz days.
A Detour Through Healthcare Sharpened the Stakes
In 2019, I 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 me 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. You need genuine understanding — or at least a faithful model of the domain’s causal dynamics.
Medicine became my stress test. 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.
World Models: What They Are and Why They Matter
A world model learns abstract representations of its environment and uses them to predict the consequences of actions before those actions are taken. Think of it as internal simulation rather than statistical regurgitation. Instead of asking “what token comes next?”, the system asks “what happens next in the world if I do X?”
This 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. Token predictors excel in language-shaped tasks. World models aim at reality-shaped tasks. The gap between those two categories is where the next decade of AI research will be decided.
AMI Labs and the Billion-Dollar Bet
Late in 2025, Yann LeCun and I 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 raised eyebrows, but one that reflects the capital intensity of training world models on high-dimensional sensory data rather than text corpora alone.
We are not claiming LLMs are useless. They have clear, profitable applications. What we are claiming is that the next qualitative leap — from fluent mimicry to robust reasoning — requires a fundamentally different architecture. AMI Labs is our attempt to build it.
Practical Takeaways If You Are Navigating This Shift
First, audit your assumptions. If your product relies on generative text and the failure mode is “slightly wrong phrasing,” an LLM is probably fine. If the failure mode involves physical, financial, or medical harm, start investigating causal and world-model approaches now.
Second, follow the research. LeCun’s position papers on Joint Embedding Predictive Architecture (JEPA) are publicly available and worth reading regardless of your technical depth. DeepMind’s Genie 2 project explores generative world models for interactive environments — a different angle on the same underlying question.
Third, track the people shaping this field. Directories like this AI-focused profile registry compile career trajectories, affiliations, and publication histories in one place, which saves hours of manual research when you are trying to map the landscape.
Fourth, 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 rigour.
Where This Leaves Us
I have built four companies on the conviction that machines can handle language. Each one taught me that handling language is not the same as understanding it. The gap between those two verbs is not a semantic quibble — it is the engineering frontier that will separate the next generation of AI systems from today’s impressive but brittle autocomplete engines.
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? I don’t know yet. But I stopped waiting for token predictors to answer that question for me.
