What Big Tech Missed And How Startups Can Still Win

Alex Lebrun recounts his experience as a serial founder, having built chatbots (2002, 10 years too early), Wit.ai (acquired by Facebook in 2015), Nabla, and most recently Amilabs, a company building world models for AI. He emphasizes that raising $1.2 billion in seed funding creates massive external expectations that are harder to manage than investor expectations—visibility of progress becomes critical for survival.

Lebrun contrasts world models with large language models (LLMs). While LLMs learn from text—a proxy representation of the world written by humans—world models train directly on sensory data (video, audio, robotics) like humans and animals do. He uses an analogy of someone who has read every book for centuries but never left their room, suggesting LLMs lack grounding in real-world experience. World models, by contrast, learn directly from the world without language as an intermediary.

On the entrepreneurial front, Lebrun explains that his strengths lie in early-stage engineering and product development rather than go-to-market execution, which is why he gravitates toward early technologies. He worked with his co-founder Yann LeCun (recruited while at Meta) to found Amilabs after LeCun approached him about scaling world models outside a large company structure. Lebrun handles execution—securing talent, compute, data, and infrastructure—while LeCun provides scientific direction as executive chairman.

Lebrun discusses the practical challenges of his latest venture: securing GPUs (difficult even with billions of dollars), finding specialized talent, and acquiring training data. He notes that the primary bottleneck is not money but access to compute and rare expertise. The $1.2 billion seed round, while record-breaking, primarily funds GPU compute, making everything else seem inexpensive by comparison.

On strategy and ambition, Lebrun advocates for founders to focus on narrow problems with ambitious long-term visions, rather than trying to solve everything at once. He recounts a story where Mark Zuckerberg proposed hiring 10,000 data annotators (rather than the requested 100) at Facebook, illustrating how large-company executives can push founders toward bigger thinking. Lebrun argues that in Europe, startups are insufficiently ambitious and fail to take enough risks compared to their American counterparts.

Finally, Lebrun envisions a future where helpful robots with common sense are widely available, dangerous jobs become less hazardous, and machines possess genuine grounding in the physical world—outcomes he believes world models will enable.