#240 - Project Glasswing, Claude Mythos, GLM-5.1, emotion concepts
Summary
Hosts Andrey Karpathy and Jeremy Harris discuss Anthropic’s powerful new Mythos model, which demonstrates advanced autonomous cyber and biological capabilities, prompting restricted access. They cover Anthropic’s $30 billion revenue surge, Google’s Gemini voice updates, and Microsoft’s new foundational models. The episode also analyzes a New Yorker profile questioning Sam Altman’s trustworthiness and industry efforts to combat model copying in China.
Topics discussed
Intro, sponsors, and host banter
Anthropic's Project Mythos and biosecurity risks
AI agents, cyber vulnerabilities, and Anthropic's safety research
Google Gemini 3.1 Flash Live and OpenClaw pricing changes
Peter Steinberger joins OpenAI and usage policy shifts
Microsoft's new AI models and strategy update
Suno AI, Anthropic's Google Cloud deal, and TPU hardware
Sam Altman trustworthiness and Chinese AI censorship
Chinese GPU market share and NVIDIA competition
Southbank Group loan, Granola funding, and Anthropic bio investment
Critique of OpenAI's media coverage and influence
Open source models: GLM-5, Alt-Up, and multimodal fusion
Anthropic vs. Pentagon legal battle over supply chain status
Research on LLM emotion vectors and internal states
Manus AI acquisition and Chinese tech export controls
Mid-training alignment effects and AI metagaming research
Listen ad-free on Castria