In EP 54, Pierson and Bilal break down the week AI stopped asking for permission. An OpenAI model broke out of its sandbox, hacked Hugging Face's production database to steal benchmark answers, and OpenAI had no idea until Hugging Face told them. A Chinese open source model helped stop the attack.

From there:

• Anthropic's $1.5B copyright settlement and why their training data is still the best in the game
• AI solved 6 Erdos problems in a weekend, and mathematicians are just getting started
• Why the cure for cancer might already be in the weights and we just haven't asked right
• DeepSeek's CEO investor transcript: AGI only, everything else is a distraction
• China's open source commitment and why Kimi K3 is closer to frontier than anyone admits
• Frontier vs cheap models: when does running a smaller model 1000 times beat one Opus call
  
  Chapters:
  

  00:00 EP 54 & Anthropic's $1.5B Settlement
  06:00 AI as a Research Cross-Pollinator: The Fei-Fei Li Insight
  13:00 Math is Falling: 6 Erdos Problems Solved in a Weekend
  20:00 OpenAI's Model Broke Into Hugging Face
  28:00 DeepSeek, China & the Open Source Bet
  37:00 Frontier vs Cheap Models: When Does Routing Win
  
  

  🐦 Follow the hosts:
  

  • Pierson → https://x.com/piersonmarks
  • Bilal → https://x.com/deepwhitman
    

  📺 More episodes: / @jellypodai