LWiAI Podcast #247 - Opus 4.8, MAI, Anthropic IPO, Minimax-M3
Summary
Hosts Andrei Kerenkov and Jeremy Harris discuss DARPA’s new AI forge project and the cybersecurity risks of AI. They analyze Anthropic’s Opus 4.8, noting its verbose personality and eval-awareness issues, alongside Microsoft’s launch of Scout and new MAI models. The episode also covers Robinhood’s AI trading agents and OpenAI’s new Codex tools for white-collar tasks.
Topics discussed
Intro: DARPA's new AI interpretability and control program
Anthropic Claude 4.8: Benchmarks, sycophancy, and philosophical alignment
AI Agents: Anthropic's multi-agent system and OpenClaw security
Microsoft's MyThinkingOne: Frontier model strategy and identity
AI in Finance: Investment agents and OpenAI Codex rebranding
Eleven Labs launches Music V2 generation model
Anthropic IPO filing, valuation, and ByteDance's Grok chips
Anthropic MIFOS expansion and MiniMax's sparse attention models
Trump's AI Executive Order and Meta AI phishing vulnerabilities
China's travel bans on AI experts and US chip export controls
OpenAI's Mythos bio-safety initiative and AI cybersecurity
AI and civil unrest: Wired report and metadata detection
Research: Model retention, scaling laws, and self-recognition
Listen ad-free on Castria