Large Language Models are fundamentally limited by their reliance on text-based data and lack of direct world experience, which prevents them from achieving true common sense or Artificial General Intelligence.
World models, which learn directly from sensory data like video, represent a vastly superior architectural path toward AGI, especially for complex, real-world applications like robotics.
Vision-Language-Action (VLA) models are an inaccurate, slow, and expensive 'hack' that misapplies LLM architecture to physical domains where it is ill-suited.
The massive cost and scarcity of GPUs for training foundational models have fundamentally altered the economics of AI startups, necessitating huge seed rounds and creating immense pressure to deliver results quickly.
Startups can still outperform large tech companies by taking significant risks on paradigm-shifting technologies that incumbents are too cautious to pursue, such as the scaling of Transformers or the development of world models.
Early 2013
Lebrun notes his early involvement in the AI space, a time when he claims most people did not even recognize '.ai' as a domain extension.
2015
Mark Zuckerberg personally initiated the acquisition of Lebrun's company, Wit.ai, via a direct email, bringing Lebrun into Facebook's AI division.
Post-2015
While at Facebook, Lebrun experienced the company's massive scale-up mentality, citing an instance where Zuckerberg escalated a proposed 100-person data collection project to 10,000 people.
Recent Past
Lebrun indicates that Yann LeCun, a key figure at Meta AI, supported the idea that world models could be developed more effectively as an autonomous startup, foreshadowing the creation of AmiLabs.
Present
Lebrun is leading AmiLabs, a company focused on building world models. He has established a globally distributed team in Paris, New York, Montreal, and Singapore, while deliberately avoiding a presence in San Francisco.
▶The Critique of Disembodied AIJul 2026
Lebrun argues that Large Language Models are fundamentally limited because they learn from text—a human representation of the world—rather than from the world itself. This lack of direct sensory experience, or 'grounding,' prevents them from developing true common sense or the ability to handle novel situations effectively.
This perspective suggests that investments in companies applying LLMs to physical, real-world problems (like robotics) may face a hard ceiling on performance, creating an opportunity for companies pursuing alternative, sensorily-grounded AI architectures.
▶The Promise of World ModelsJul 2026
He champions 'world models' as the superior path toward more capable AI. These models learn directly from raw sensory data like video and audio, building an intuitive understanding of physics and causality without the intermediary of human language, which he believes is essential for creating genuinely helpful robots.
Lebrun is positioning his company, AmiLabs, as a high-risk, high-reward bet on a potential paradigm shift in AI, moving from a text-centric to a perception-centric approach to building intelligence.
▶The New Economics of AI Foundational ModelsJul 2026
Lebrun highlights how the need for massive-scale GPU clusters has fundamentally altered startup financing. The requirement to purchase thousands of GPUs, costing 'billions of euros,' necessitates exceptionally large seed rounds and creates immense external pressure to show results within a short timeframe.
The capital-intensive nature of building foundational models creates a significant barrier to entry, favoring extremely well-funded teams and potentially limiting the competitive landscape to a few major players, even within the startup ecosystem.
▶Startup Strategy in the Age of Big Tech AI
According to Lebrun, a startup's primary advantage over incumbents is its ability to take significant risks on unproven, paradigm-shifting ideas that large companies often ignore. He advises founders to solve a narrow problem while articulating an ambitious long-term vision, citing OpenAI's success with scaling Transformers as an example.
This framework suggests that the most disruptive AI companies will not be those making incremental improvements on existing models, but those challenging the core architectural assumptions of the current leaders.