#213 - Midjourney video, Gemini 2.5 Flash-Lite, LiveCodeBench Pro
Summary
This episode covers Midjourney’s new video generation tool, Google’s Gemini 2.5 model updates, and YouTube’s integration of Veo3. It also discusses OpenAI dropping Scale AI, Zoox’s robo-taxi production, and new benchmarks for coding and model abstention. Finally, it touches on safety research regarding jailbreaks and emergent misalignment, plus a major OpenAI defense contract.
Topics discussed
Sponsors: Box and Outshift
Intro and guest background
Midjourney video generation launch
Google Gemini 2.5 Pro Flash model
Apple Intelligence voice mode
YouTube Shorts adds Veo3 AI
OpenAI ethics documentation site
Meta hires Scale AI leadership
Zoox self-driving vehicle production
Live Code benchmark for coding agents
LLM abstention benchmark study
Minimax open-source reasoning model
Waymo driving laws and data scaling
Jailbreak suffixes and attention hijacking
Misalignment features in LLMs
OpenAI and DoD partnership
Listen ad-free on Castria