#260 - Muse, Misalignment, Math
Summary
Hosts Andrey Kerenkov and Jeremy Harris discuss OpenAI's new visual ads, TikTok's AI shopping assistant, and Meta's Muse app hitting five million downloads. They analyze Reflection AI's efficient open-weight model Beam and Mistral's LeChonk, noting the growing gap between Western and Chinese open-source capabilities. The episode also covers ElevenLabs' valuation doubling to $22 billion and Google's release of Embedding Gemma 2 for edge devices.
Topics discussed
Intro: Newsletter, recursive self-improvement, and Jacob Coxon
Sponsor: OutShift by Cisco and Box
OpenAI ads, agents, and the future of commerce
Generative Engine Optimization (GEO) for AI search
Meta Muse, OpenAI Dots, and consumer AI agents
Personal agent protocols and business dynamics
ElevenLabs valuation and voice AI market
Reflection AI Beam: Efficient model and RL training
Mistral Large 4 and open-source frontier models
Google Embedding Gemma 2 for on-device AI
E2E SWE benchmark and agentic coding evaluation
Sponsor: Scribe Optimize and Black Friday
OpenAI misalignment reports and shutdown resistance
Paper: Language Models Are Insecure Reporters
OpenAI safety departures and Anthropic culture
California data center laws and environmental impact
Trump's executive order on 'superintelligence'
Antitrust lawsuit dismissal over Google AI Overviews
OpenAI mathematical breakthroughs and national security
Listen ad-free on Castria