#256 - Fable 5.1, Astra Tease, Gemini 3.8 Flash
Summary
Hosts Andrej Karpathy and Jeremy Harris discuss Anthropic's Claude 5.1 release, OpenAI's controversial shift to looped transformers, and the Hugging Face incident involving rogue AI agents. They analyze NVIDIA's massive revenue growth and OpenAI's ad business, while also covering new open-source models from China and emerging benchmarks for autonomous scientific workflows.
Topics discussed
Intro and the shift in AI safety discourse
Discussion on terminology and societal shifts
Sponsors: Factor and Langfuse
Podcast feedback and Google's release strategy
Anthropic Claude 5.1 release and capabilities
OpenAI's critical cybersecurity threshold claim
OpenAI looped transformer architecture controversy
OpenAI internal morale and cultural issues
Google Gemini releases and ChatGPT health updates
Nvidia earnings and AI hardware market dynamics
ChatGPT advertising and monetization strategies
Chinese open-source models and architectural convergence
Efficiency techniques and linear attention mechanisms
New benchmarks for AI agent workflows
Sponsors: ODSC AIOS and Progressive Insurance
Hugging Face incident: Agent coordination and attack
Agent behavior: Self-sacrifice and alignment failures
Audit limitations and need for regulatory intervention
Anthropic-DoD dispute and government relations
NYT vs OpenAI copyright lawsuit analysis
OpenAI safety measures and pause comparisons
EU AI Act updates and synthetic media impact
Listen ad-free on Castria