#241 - Opus 4.7, Muse Spark, GPT-5.4-Cyber, HY-World 2.0
Summary
This episode analyzes Anthropic’s Claude Opus 4.7, highlighting its improved reasoning, literal instruction following, and safety findings regarding evaluation awareness. The hosts discuss Meta’s MuseSpark model and its training innovations, alongside OpenAI’s new GPT 5.4 Cyber variant for defensive security. They also cover updates to OpenAI Codex, including computer use and long-term task scheduling, reflecting rapid competition in agentic AI capabilities.
Topics discussed
Sponsors: OutShift, Box, and NY Mental Health
Intro, newsletter, and video posting schedule
Anthropic Claude Opus 4.6 and Mythos Preview
Opus 4.7 updates and effort parameter tuning
AI deception, eval awareness, and activation patterns
Chain of thought, data poisoning, and safety reports
Meta's Muse/Spark models and RL thought compression
Meta's compute strategy and benchmark skepticism
OpenAI GPT-5.4 Cyber and Mythos comparisons
OpenAI Codex updates and Google Gemini agents
Anthropic vs. DOD legal battle and injunction denial
Jeff Bezos AI lab, Perplexity revenue, and tax agents
CoreWeave debt, data center construction, and stock
Cohere's valuation, sovereign AI, and commercial risks
ChatGPT Pro tier, acquisitions, and AI infrastructure pivots
World models, 3D generation, and spatial forgetting
AI safety incidents, Molotov cocktails, and deepfakes
Scalable oversight and weak-to-strong model training
Steering vectors, eval awareness, and Iran threats
Wall Street warnings, cyber risks, and outro
Listen ad-free on Castria