Subquadratic has now shared more details about its new model. But some are still skeptical.

OpenAI researchers show that reinforcement learning on desired behavioral traits like truthfulness and corrigibility works across domains. Training on health data also improved deception detection, and the model scored better on 44 out of 53 benchmarks. The approach differs from Anthropic's constitution-based method.

Norway is banning generative AI tools in elementary schools starting in late August. Students in grades 1 through 7 won't be allowed to use AI at all; secondary schools will permit it only under supervision. Prime Minister Stoere says children must first "learn to read, write, and do math."

Nobel Prize winner John Jumper is leaving Google Deepmind for Anthropic after nearly nine years. Days earlier, Gemini co-lead Noam Shazeer left for OpenAI. Weeks before that, AlphaGo researcher David Silver started his own company. Three of Google's most prominent AI minds, gone within months.