Elon Says 100X, Memory Prices Spike 500%, Unitree Robot Outruns Usain Bolt with Emad Mostaque | #282

In this talk

1. 8:52

   AI safety and frontier model access

   The panel examines OpenAI’s pause on some reinforcement learning training. Emad sees both governance concerns and public relations motives, while the group debates whether labs are keeping their most capable models for internal work and how far those models are ahead of public releases.

2. 27:32

   Model gains, convergence, and agent teams

   The panel discusses claims of 100-fold gains in intelligence density and further efficiency from specialized or sparse models. They also debate whether leading models are converging on similar reasoning, and whether a shared architecture could create common blind spots. Peter and Dave suggest that teams with thousands of AI agents could tackle large projects, but people need to decide what goals to assign them.

3. 1:26:08

   Memory constraints and faster robots

   Peter says memory prices have risen fivefold in a year and argues that memory supply is becoming a key constraint as AI workloads expand. The panel discusses high-bandwidth memory, alternatives such as etching model weights into chips, and Unitree’s robot reaching 12.66 meters per second. Emad questions the usefulness of human-shaped robots, while the panel considers safety rules for powerful machines.

4. 1:46:09

   AI applications in delivery and medicine

   The panel covers Zipline and Uber’s plan for up to one million autonomous drone deliveries per day, with Uber acting as a platform for third-party services. They also discuss a personalized mRNA cancer vaccine that showed reduced recurrence or death in a melanoma trial, and cell simulation as a way to prioritize experiments before running them in a lab.

5. 59:08

   AI governance and access

   The panel debates Anthropic’s reported plans for founder supervoting shares and the role of its Long-Term Benefit Trust. They also discuss whether frontier AI regulation could favor major labs, and Emad predicts that limited GPU and memory capacity may lead to calls for a universal right to AI access.