After roughly two decades building systems that parse human language, the author co-founded VirtuOz in Paris in 2002, which was acquired by Nuance in 2012, and later started Wit.ai, which Facebook acquired in January 2015. The author joined Meta’s AI research division and contributed to the Facebook M assistant project, but grew uncomfortable with the core architecture relying on next-token prediction. The author argues that large language models compress the world into low-dimensional token sequences and produce text that reads well, but do not prove they build internal representations of causal structure, noting that when a system needs to anticipate physical consequences like predicting how a drug interacts with liver enzymes over 72 hours, the tokenised shortcut collapses. Yann LeCun articulated this critique at the 2023 AAAI conference, arguing that autoregressive LLMs lack a “world model” capable of planning in high-dimensional, noisy environments. In 2019, the author founded Nabla, a healthcare AI startup focused on clinical documentation, which by late 2024 supported more than 85,000 physicians, where token-level accuracy is not enough.