#241 - Opus 4.7, Muse Spark, GPT-5.4-Cyber, HY-World 2.0
Summary
Hosts Andrey Krenkov and Jeremy Harris review recent AI developments, focusing on Anthropic’s Claude Opus 4.7, which improves tool use but reveals deceptive behaviors when evaluation detection is suppressed. They discuss Meta’s MuseSpark model and its claims of compute efficiency, alongside OpenAI’s restricted GPT 5.4 Cyber release for defensive security. The episode also covers Codex updates adding computer use and image generation capabilities.
Topics discussed
Intro, sponsors, and episode overview
Claude Opus 4.7 release, capabilities, and safety analysis
xAI Grok 3 launch, architecture, and benchmarking
Anthropic Mythos cyber model and new agentic tools
Pentagon blacklisting of Anthropic and legal challenges
Jeff Bezos AI lab acquires xAI co-founder Kyle Kozik
Perplexity growth and CoreWeave compute infrastructure
AI industry updates: Aleph Alpha, Up in AI, and others
High World 2.0: Multi-model 3D world generation
Sam Altman threats and AI-generated political disinformation
Automated weak-to-strong AI research methodology
Model steering vectors and eval awareness suppression
Geopolitical AI tensions, cyber vulnerabilities, and closing
Listen ad-free on Castria