What Big Tech Missed And How Startups Can Still Win
Alexandre LeBrun, CEO of AMI Labs, discusses his career building and selling AI companies, emphasizing his strategy of tackling problems "20 years too early." He delves into AMI Labs' contrarian bet on "world models" over traditional LLMs, highlighting their ability to learn directly from real-world sensory data, unlike LLMs which learn from human-written text. LeBrun explains how this approach is critical for developing intelligent robots and avoiding the pitfalls of Vision-Language Assistants (VLAs). He also touches upon the challenges of securing talent, data, and compute for such an ambitious project, the strategic choice of location, and the importance of holding an extremely large vision while solving a narrow problem for early founders.
Alexandre LeBrun (Alex), a serial entrepreneur with four startups under his belt, including Wit.ai (sold to Facebook), discusses his journey and the foundational principles behind his latest venture, AMI Labs. He reflects on his tendency to be “20 years too early” with his companies, starting with a chatbot company, Vertos, in 2002 when the market wasn’t ready. This early experience taught him that his strength lies in taking very early-stage technology, conceptualizing a product around it, and understanding nascent markets, rather than excelling in established go-to-market strategies.
The Contrarian Bet: World Models vs. Large Language Models (LLMs)
AMI Labs is built on the ambitious idea of developing “world models,” a fundamentally different approach to AI compared to mainstream LLMs.
- Limitations of LLMs: Alex argues that LLMs learn from human-written text—a proxy for the real world. They are akin to someone who has read all books but never experienced the world directly, thus lacking common sense and the ability to generalize to novel situations. While excellent for “language-first” tasks, mathematics, and programming (which are low-dimensional, discrete symbol problems), LLMs are not designed for the complexities of the physical world.
- World Models’ Approach: In contrast, world models learn directly from sensory data like video, audio, and robotics-derived touch data, mimicking how humans and animals learn “from scratch like a baby, without any language.” This grounding in real-world experience is crucial for true intelligence.
Applications and Why World Models are Necessary
- Robotics: A primary application. Current robots are highly specialized and perform well in controlled, narrow tasks. However, their “brains” are primitive, making them unsafe and useless in open, dynamic environments (e.g., homes, streets). World models aim to equip robots with “actual brains” capable of common sense and robust interaction with the physical world.
- Critique of VLAs (Vision-Language Assistants): Alex views VLAs as a “very bad hack” that misuses LLMs. They are inaccurate, slow, and computationally expensive for real-time, high-dimensional tasks like robotics, where low latency and cost-efficiency are paramount.
- Current State: While world models are costly to train (comparable to LLMs), their inference is expected to be more lightweight. Some ideas derived from world models (e.g., Jepa) are already in production in fields like autonomous driving (e.g., Waymo, Wave).
Challenges and the Founding of AMI Labs
- Risk-Taking: Alex highlights that startups like AMI Labs can take risks that large corporations cannot. He recounts an anecdote at Meta where a chatbot generated inappropriate content, illustrating the legal and brand constraints that hinder radical innovation in big companies, allowing startups (like early OpenAI) to push boundaries.
- Bottlenecks: Building world models requires immense talent (many of whom are currently focused on LLMs), vast amounts of data, and significant GPU compute, which is difficult to secure even with substantial funding.
- Collaboration with Yann LeCun: AMI Labs was co-founded with Yann LeCun, who serves as the Executive Chairman, steering the scientific vision. Alex, as CEO, manages operations, balancing research freedom with strategic direction to ensure product delivery.
- Record Funding and Expectations: AMI Labs secured Europe’s largest seed round (€1 billion), driven by the astronomical costs of compute. Alex emphasizes that the “real cost” of such funding is the heightened expectations from investors and the public. Failure to show progress within a few years makes survival challenging.
- Strategic Location: AMI Labs operates from Paris, New York, Montreal, and Singapore, intentionally avoiding a primary SF base to attract deeply committed talent.
Advice for Founders and Future Vision
Alex encourages founders, particularly in Europe, to embrace greater ambition and risk-taking. He shares a story where Mark Zuckerberg, when asked to approve hiring 100 people for Wit.ai, responded with a suggestion to hire 10,000, underscoring a different scale of ambition.
His key advice for early-stage founders is to “go narrow, aim massive”: choose a very specific problem but harbor an extremely large, long-term vision within that niche. This approach allows for credibility while maintaining ambitious goals. He envisions a future where AMI Labs’ success leads to helpful, common-sense robots that make dangerous or difficult jobs safer and easier, significantly transforming daily life.