A Lab Stops Its Own Biggest Run

Hosts: Lenar Kess, Damra Vol. OpenAI says it has paused reinforcement learning training on its latest deployment-bound models while new security and monitoring requirements catch up. On the same day, an outside group graded every frontier lab on whether its control practices are actually implemented, Anthropic disclosed three unreleased internal models, and a governor signed data center siting rules while Nvidia backed a 4.25 gigawatt build next door.OpenAI announced a temporary pause on reinforcement learning runs for models intended for deployment — the pause is specific to that phase, not to all training.Pacing model development in an era of cyber-critical capabilities is the policy post underneath the announcement, and it names cyber capability as the pacing constraint.Greg Brockman and Sam Altman gave two different accounts of the same decision — confidence setting the pace versus capability outstripping safety.Max Zeff reports the next frontier model, Astra, is held up on security and alignment work, and the largest planned run remains on hold.Ethan Mollick flagged the disclosed figure of roughly 20% of research inference compute going to chain-of-thought monitoring.Steven Adler's Guidelight scorecard grades frontier labs on control practices and finds all of them at most partially implemented, on public evidence alone.Anthropic's protein design campaign reports a 35% binder success rate, and Ravid Shwartz Ziv accounts for the 30k-token expert prompt and roughly 12,500 H100-hours behind it.Governor Josh Shapiro signed data center standards for Pennsylvania the same day a 4.25 gigawatt Ohio build was reported at 105 billion dollars.Cerebras CS-4 and Etched's 700 million dollar raise arrive as average token prices halve.LangSmith Tuned Evaluators and fx, a 6.3 mebibyte coding agent in Zig, both aim at agent work from opposite ends of the size range.

Ver todos los episodios