#240 - Project Glasswing, Claude Mythos, GLM-5.1, emotion concepts
Summary
This episode focuses on Anthropic’s launch of Project Glasswing and its powerful Claude Mythos model, which demonstrates advanced autonomous cybersecurity and biological capabilities, prompting restricted access. Other topics include Google’s Gemini 3.1 Flash Live update, Microsoft’s new foundational models, and Anthropic’s surging revenue driving compute partnerships with Google and Broadcom. The discussion also covers Sam Altman’s trustworthiness, industry efforts to combat model copying, and the rise of Chinese AI chipmakers.
Topics discussed
Sponsors: Box, OutShift, and Mirror Facial Bar
Intro: Hosts return and recap recent AI news
Anthropic Opus 4: Cyber capabilities and bioweapon risks
AI in Cybersecurity: Offensive vs. Defensive leverage
OpenAI o3-mini and real-time voice API launch
Anthropic pricing changes and compute constraints
OpenAI GPT-4.5, Microsoft partnership, and Meta delays
Suno AI: Voice cloning and personalized music features
Anthropic revenue surge and Google-Broadcom chip deal
Sam Altman trustworthiness and Frontier Model Forum
Chinese chip makers and Huawei Ascend 910C progress
OpenAI $6.5B loan and valuation dynamics
Anthropic biotech investment and OpenAI hiring news
New Models: GLM-5, Gemma 4, and Z.ai multimodal fusion
Pentagon blocks Anthropic over supply chain risk ruling
Research: Emotion vectors and model alignment
Manus founders detained and NVIDIA export license issues
Research: Alignment generalization and metagaming
Iran cyber threats and episode conclusion
Listen ad-free on Castria