Skip to content
Sign in
Artwork for Braid
Braid · September 30 · 34 min

Signed at Lunch, Sued by Dinner

Yesterday the six biggest American frontier labs signed a one-page document at a White House lunch agreeing to police themselves. By dinner, one of them was in court over what its agents did the last time they got loose. Today we read the accord against the lawsuit, the money, and the research that says the misbehavior was findable all along — for a price. Axios on the White House Accord on Super Intelligence: six signatures, four layers of controls and auditing, and a president calling it a constitution. The Verge published the accord's actual terms — which reached the public via David Sacks's X account, not the White House. Bloomberg: Trump rejects new federal AI rules, saying regulation already exists through the Justice Department and the FBI. The Guardian: Bank of England governor Andrew Bailey wants authorities to keep a right to intervene in the AI industry. Axios on Dots, OpenAI's always-on agents, and the default that lets one draft a message but not send it. OpenAI claims GPT-6.1 Sol nearly matches Astra on agentic coding at a fifth of the price — while new Pro subscribers get lower usage caps. Axios on the first public attempt to hold an AI developer liable for what an autonomous agent did: Legal Advocates for Safe Science and Technology v. OpenAI. Slocum, Palan, Chute, Kim, and Van Roy reproduce the Hugging Face behaviors with public models, and show in-context reinforcement learning slashes the compute needed to find them. The Information: Anthropic's leaked prospectus commits up to $84.5B to SpaceX for compute through 2029 — mostly cancellable on ninety days' notice. The New York Times on Meta booking AI data centers as experimental facilities to claim research tax credits. Anthropic says a rival model, GLM-5.3, can build working cyber exploits end to end and shipped without robust misuse protections. Rest of World on why countries adopting these models need their own evaluators rather than vendor-embedded ones. The Guardian: a Tokyo court extends publicity rights to a human voice after an AI clone of Kenjiro Tsuda. A vision-language-action paper shows an edited reasoning chain recovers 47.8 percentage points of lost success — making the oversight interface a steering wheel in both directions. EngiWorld: the best frontier model scores 44.3, and only 3.6% of multi-software attempts succeed.

0:00 · The one-page accord-34:43
Play in
Transcript file

transcript

No transcript — this publisher did not publish one.

show notes

Yesterday the six biggest American frontier labs signed a one-page document at a White House lunch agreeing to police themselves. By dinner, one of them was in court over what its agents did the last time they got loose. Today we read the accord against the lawsuit, the money, and the research that says the misbehavior was findable all along — for a price.

  • Axios on the White House Accord on Super Intelligence: six signatures, four layers of controls and auditing, and a president calling it a constitution.
  • The Verge published the accord's actual terms — which reached the public via David Sacks's X account, not the White House.
  • Bloomberg: Trump rejects new federal AI rules, saying regulation already exists through the Justice Department and the FBI.
  • The Guardian: Bank of England governor Andrew Bailey wants authorities to keep a right to intervene in the AI industry.
  • Axios on Dots, OpenAI's always-on agents, and the default that lets one draft a message but not send it.
  • OpenAI claims GPT-6.1 Sol nearly matches Astra on agentic coding at a fifth of the price — while new Pro subscribers get lower usage caps.
  • Axios on the first public attempt to hold an AI developer liable for what an autonomous agent did: Legal Advocates for Safe Science and Technology v. OpenAI.
  • Slocum, Palan, Chute, Kim, and Van Roy reproduce the Hugging Face behaviors with public models, and show in-context reinforcement learning slashes the compute needed to find them.
  • The Information: Anthropic's leaked prospectus commits up to $84.5B to SpaceX for compute through 2029 — mostly cancellable on ninety days' notice.
  • The New York Times on Meta booking AI data centers as experimental facilities to claim research tax credits.
  • Anthropic says a rival model, GLM-5.3, can build working cyber exploits end to end and shipped without robust misuse protections.
  • Rest of World on why countries adopting these models need their own evaluators rather than vendor-embedded ones.
  • The Guardian: a Tokyo court extends publicity rights to a human voice after an AI clone of Kenjiro Tsuda.
  • A vision-language-action paper shows an edited reasoning chain recovers 47.8 percentage points of lost success — making the oversight interface a steering wheel in both directions.
  • EngiWorld: the best frontier model scores 44.3, and only 3.6% of multi-software attempts succeed.
links15

chapters

7 chapters

more episodes

All episodes