#247 - Opus 4.8, MAI, Anthropic IPO, Minimax-M3
Summary
Hosts Andrey Kerenkov and Jeremy Harris discuss Anthropic’s Opus 4.8, highlighting its verbose personality, eval-awareness risks, and new dynamic workflows. They analyze Microsoft’s launch of the Scout agent and its MyThinkingOne model, emphasizing enterprise security and the strategic positioning of second-tier AI labs. The episode also covers Robinhood’s new AI trading agents and OpenAI’s expanded Codex tools for white-collar productivity.
Topics discussed
Intro, DARPA guest plug, and Progressive ad
Anthropic Claude 4.8 release, eval awareness, and corrigibility
Microsoft Copilot Studio and OpenClaw agent framework
Microsoft's custom LLM strategy and tier-2 lab dynamics
Robinhood AI trading agents and OpenAI Codex branding
Anthropic IPO filing and trillion-dollar valuation implications
ByteDance's LPU chips and bypassing US export controls
Google CapEx, Anthropic MIFOS expansion, and Cognition funding
Open source models: DeepSeek V4 and MiniMax performance
US AI safety executive order and Trump administration policy
AI social engineering attacks and US travel restrictions for researchers
Clarification of AI chip export controls and loopholes
OpenAI biosecurity initiative and Mythos vulnerability findings
US spy agencies acquiring AI chips and domestic extremism concerns
Research papers on model learning, entropy, and outro
Listen ad-free on Castria