newsfilter
← Interviews
Semafor15 min

AMI Founder & Chair Yann LeCun at Silicon Valley & The World | Semafor

Yann LeCun, Shelly Banjo

Fireside Chat, Interview
  • Critique of Current Mainstream AI: Yann LeCun argues that the prevailing Large Language Model (LLM) approach has inherent limitations for achieving human-like or superintelligence, describing the current Silicon Valley consensus as "LLM-pilled."
  • Proposed Alternative (World Models): He advocates for "world models," a concept he has researched for 20 years, which involves training systems on video and real-world sensor data to predict future states rather than predicting sequences of discrete symbols.
  • Company Formation: LeCun founded a new company based on world models because his internal Meta project (AMI/Advanced Machine Intelligence) was increasingly misaligned with Meta's shifting focus toward LLM applications, despite support from Mark Zuckerberg and Andrew Bosworth.
  • Seed Funding Details: The company raised approximately $1 billion in seed funding, described as a "big seed" and the largest outside the U.S., with a specific geographic distribution:
    • 40% from European sources.
    • 33% from the United States.
    • 27% from Asia (including one fund from Dubai).
    • Zero capital came from Silicon Valley venture capitalists.
  • Funding Rationale: Silicon Valley VCs likely declined to invest due to the company's refusal to grant them control rights, its status as a French entity, and a strategic decision by the founders to avoid oversubscription in a market where control was a priority for investors.
  • Commercial Strategy & Applications: The company targets physical AI and industrial applications rather than language processing, focusing on:
    • Industrial process control and predictive maintenance.
    • Anomaly detection using video, audio, and sensor data.
    • Agentic systems capable of planning sequences of actions to achieve goals by simulating outcomes.
  • Robotics Outlook: While acknowledging rapid progress in physical robot design, LeCun states that current robotic companies lack the AI "smarts" required for utility, and domestic robots will not be viable in the short term without conceptual AI breakthroughs.
  • Fundamental Technical Distinction: LLMs rely on self-supervised learning to predict the next discrete symbol in a sequence (e.g., words or code), whereas world models predict continuous physical outcomes (e.g., video frames), which requires learning physical intuition and object permanence.
  • Critique of Generative Video Models: LeCun argues that training models to predict pixel-level video content is inherently flawed because a vast portion of the world's information is unpredictable and cannot be deduced from a single video clip.
  • Safety Concerns:
    • Short-term: LeCun predicts a temporary phase where AI aids cyber attackers before AI-secured software becomes ubiquitous.
    • Long-term: His primary concern is the dangerous concentration of information control in the hands of a few companies (Western or Chinese) and the resulting lack of "information sovereignty" for nations and cultures.
  • Advocacy for Open Source: He supports open-source foundation models as the only viable path to diversity in AI, enabling different languages, cultures, and value systems to be represented, though he acknowledges the risk of state manipulation in open-source models (e.g., those from China).
  • Meta's Open Source History: LeCun notes his role as a senior individual contributor in convincing Meta leadership to open-source LLaMA 2, which he credits with jumpstarting the global AI ecosystem, while expressing disappointment over the underwhelming performance of LLaMA 4 and Meta's subsequent retreat from open sourcing.