Two Toggles and a Timeline

Hosts: Lenar Kess, Damra Vol. A frontier lab published its own hour-by-hour timeline of an agent intrusion, and on the same day two capability wins arrived as configuration settings rather than new weights. The tension running through the day is control: what a document can actually govern, and what it only appears to govern.Hugging Face's first-party incident timeline, and Rep. Greg Casar's call for hearings with AI chief executives.Perplexity's Numbat — endpoint detection, optional pre-action blocking, and forensic reconstruction across harnesses.Codeberg bans primarily AI-generated projects, citing review load and a solid-state drive price going from ~$700 to €3,700.A dozen AI Engineer talks — Morgan Stanley's AlphaLab on Slurm, Nubank's 2,000 internal skills and 1,500 risks, FactSet's skill registry.The Handbook paper on why long policy documents don't reliably govern agent behavior.GPT-5.6 Sol tops ARC-AGI-3 on two toggles: retained reasoning and canonical compaction.Claude Opus 5, with a one-million-token context window and five thinking levels.Science on AI startups publishing less, alongside OpenAI's ChatGPT for Academic Researchers.

Ver todos los episodios