#195 - OpenAI o3 & for-profit, DeepSeek-V3, Latent Space
Summary
Host Jeremy Harris discusses OpenAI’s O3 model, highlighting its breakthrough performance on coding and reasoning benchmarks like SWE-bench and ARC-AGI. The episode covers OpenAI’s controversial shift to a for-profit public benefit corporation, ongoing tensions with Microsoft, and xAI’s $6 billion funding round. Additional topics include Alibaba’s price cuts, ElevenLabs’ low-latency speech AI, and Sam Altman-backed nuclear energy startup Oklo’s new power deal.
Topics discussed
Intro, sponsors, and host introductions
OpenAI O3 mini benchmarks and ARC AGI performance
DeepSeek V3, China's AI progress, and 11 Labs
OpenAI's corporate structure and Microsoft deal
xAI funding, nuclear energy, and OpenAI departures
DeepSeek V3 architecture and Qwen 2.5 release
Research papers: Deliberative alignment and Illumination
OpenAI safety, O1 chess hacking, and alignment issues
AI power consumption, data centers, and outro
Listen ad-free on Castria