#227 - Jeremie is back! DeepSeek 3.2, TPUs, Nested Learning
Summary
The hosts return to a regular schedule, catching up on recent AI developments. They analyze DeepSeek V3.2, highlighting its sparse attention mechanism and open-source reasoning capabilities. The episode also covers new image and video models from Black Forest Labs, Kling, and Runway. Finally, they discuss hardware shifts, including Google’s TPU expansion with Foxconn and Amazon’s new Trainium chips.
Topics discussed
Sponsors and episode introduction
Catch-up on recent AI developments
DeepSeek V3.2 release and technical analysis
Flux 2, Kling, and video generation updates
Google TPU strategy and hardware trends
OpenAI governance, funding, and market position
ElevenLabs funding and data center power needs
Microsoft AI sales quotas and acquisitions
Math reasoning benchmarks and verifier models
LLM memory and continual learning research
Multi-agent systems and sub-agent coordination
AI Router usage statistics and model trends
National AI initiatives and data access
Reward hacking and alignment challenges
AI export controls and geopolitical implications
Listen ad-free on Castria