Fireside Chat, Interview
What Big Tech Missed And How Startups Can Still Win
- Alex K. stated the primary cost of raising 1.2 billion (approx. 1 billion EUR) is not dilution but the creation of high external expectations; failure to deliver visible progress within two years could make the company unsustainable.
- Investors are aligned with the long-term vision, but public perception requires tangible output to maintain survival.
- Alex K. and his co-founder Yan LeCun (Executive Chairman) previously worked together at Meta and Nabla before founding Ami Labs.
- They met at Y Combinator; Alex K. was founding Algolia while LeCun founded Wit.ai.
- Ami Labs is raising approximately 1 billion EUR in seed funding to build foundational "world models" rather than applying existing large language models (LLMs).
- This is an "upstream" shift from applied AI to building the core models themselves.
- The cost structure is driven by the need to purchase thousands of GPUs, necessitating billions in capital.
- The company distinguishes "world models" from LLMs by rejecting text-only training data in favor of direct sensory experience.
- LLMs are described as learning from proxy texts written by humans, lacking direct physical experience.
- World models train on video, audio, and robotics data to simulate how babies and animals learn from the real world.
- This approach aims to provide the "common sense" and grounding missing in current LLMs.
- Ami Labs targets robotics applications where current "vision-language-action" (VLA) models fail.
- VLA is characterized as a slow, expensive hack that lacks the real-time latency and accuracy required for home robots.
- Current robots are too narrow in scope and lack the safety to operate in open environments without world models.
- The company maintains a non-traditional geographic strategy, operating primarily in Paris with hubs in New York, Montreal, and Singapore, while deliberately avoiding a physical San Francisco office.
- The decision prioritizes team commitment over ecosystem proximity, requiring talent to relocate rather than allowing remote flexibility.
- The team remains connected to the SF ecosystem through frequent travel but avoids the high churn of temporary setups.
- Alex K. emphasizes that founders should aim for "narrow problems with big visions" rather than trying to solve everything immediately.
- He cites his 2002 chatbot attempt as a failure due to being "20 years too early" and lacking market readiness.
- He recounts a story where Mark Zuckerberg initially approved a plan for 100 concierge workers, then jokingly suggested 10,000, illustrating the need to temper ambition with realistic risk assessment.
- Current bottlenecks for the project are securing talent, data, and compute (GPUs).
- While data and talent are accessible, securing sufficient GPU compute is described as increasingly difficult even with capital.
- The belief that LLMs will not lead to AGI has shifted from a contrarian view to a widely accepted premise among experts.
- The company's split responsibilities involve LeCun leading scientific direction and Alex K. handling execution, resource management (electricity, GPUs, office), and balancing researcher autonomy with product delivery.
- Managing researchers is described as a "tightrope walk" between giving too much freedom (random results) and too much direction (fleeing talent).
- Alex K. advises early-stage founders to take significant risks and be ambitious, noting that European startups often suffer from under-ambition.
- He argues that high-risk bets are often necessary to solve hard problems, even if they fail.
- The long-term vision for Ami Labs includes the deployment of helpful, general-purpose robots that perform dangerous or difficult tasks, granting machines a form of common sense.