newsfilter
← Interviews
Semafor15 min

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

Yann LeCun, Shelly Banjo

Fireside Chat, Interview
  • Progress toward human-like or super intelligence requires abandoning current LLM approaches in favor of world models, which predict action outcomes to enable agentic systems, as LLMs are limited by their inability to model physical reality and blind spots regarding consequences.
  • Significant advancements in world models were achieved over the last five years through self-supervised learning from real-world video, addressing a two-decade struggle where progress was previously stalled for 15 years.
  • A new company was established to accelerate industrial applications such as physical AI, process control, predictive maintenance, and anomaly detection, focusing on video, audio, and sensor data rather than language to complement Meta's internal work and LLM-focused AI organization.
  • Robotics is identified as a major business sector, yet domestic robots are not expected to be viable in the immediate future due to current intelligence deficits, necessitating specific conceptual breakthroughs before practical utility can be achieved.
  • Training systems to predict video content at the pixel level is deemed unsolvable due to the unpredictability of most world information, contrasting with the approach of predicting action sequences.
  • The near-term AI landscape will see a brief period where tools facilitate cyber attacks by favoring attackers, followed by an era of unprecedented software safety as AI hardens future software production.
  • Concerns regarding the extreme danger of power concentration in a few companies controlling information sources drive a call for high diversity in AI assistance, similar to the press, which proprietary systems currently fail to provide.
  • Open source foundation models are positioned as the sole path for AI sovereignty and culturally diverse assistants, with the best current models originating from China, despite risks of regional manipulation that open sourcing can partially mitigate through unconditioning.
  • The open sourcing of Llama 2 previously jumpstarted the global AI ecosystem, while recent disappointment with Llama 4 has triggered a reboot of the entire AI effort by Meta leadership.