Friday, 25 September 2026
Three Labs Are Reportedly Founding Their Own Regulator — and Want an Anti-Regulator to Run It
Google, OpenAI and Anthropic Are Reportedly Building Their Own Regulator — and Want an Anti-Regulator to Run It
The Information reported yesterday that the three largest frontier labs have agreed to stand up a shared self-regulatory body, provisionally the Standards Authority for Frontier AI, modelled on FINRA: common benchmarks, third-party testing before deployment, incident-reporting guidelines, and a written form for the voluntary safety commitments everyone currently interprets privately. As a self-regulating organisation it would need no act of Congress and no signature from the White House. The idea is Demis Hassabis’s, from a personal essay on July 14. The names being approached are the part worth sitting with: Condoleezza Rice, former Biden technology official Arati Prabhakar, the venture capitalist David Friedberg, Metr’s Beth Barnes and Paul Christiano as scientific advisers — and, as lead candidate for chief executive, Sriram Krishnan, who left the White House AI job in June having said on the record that “there will not be an FDA for AI.” Target launch is late this year or early 2027. Read it carefully, though: this is investigative reporting, not a joint announcement. What the companies have confirmed on the record is narrower — OpenAI’s Chris Lehane said on September 15 that the three had been working together on safety for several weeks and saw no need for an antitrust waiver to do it. The distance between “working together for several weeks” and “agreed to found a regulator” is exactly the distance The Information is reporting across, and none of the three has closed it publicly.
Amazon Will Let an Agent Run Your Shop. It Will Not Let One Buy Anything.
At Amazon Accelerate on Tuesday, Amazon opened its selling-partner tools to outside AI agents: a beta plugin that lets a seller check inventory, move prices, edit listings and pull analytics from inside Anthropic’s Claude or Amazon’s own Quick assistant, without ever loading Seller Central. Every action still stops for human approval and leaves an audit trail. Mary Beth Westmoreland, VP of worldwide selling partner experience, put the goal plainly: “Our vision was that they would never have to log into Seller Central. We would just bring it to them where they work.” Amazon says over 90% of its sellers already use outside AI somewhere in their operation, which reads less like a boast than an explanation. The interesting half is the half that stayed shut. Sellers get an open door; shoppers don’t — the consumer storefront remains closed to third-party agents, and Amazon blocked Meta’s shopping agent Muse only recently. The side of the marketplace that pays Amazon $46.8 billion a year in fees gets to automate. The side that spends money still has to come to the website.
Source: GeekWire, September 23, 2026
Google Is Putting Four TPUs on a SpaceX Rocket Next Thursday to See What Breaks
Project Suncatcher’s first satellite — MVP, roughly the size of a refrigerator, built with Planet Labs — flies October 1 on a Falcon 9. On board: four custom TPUs, about a kilowatt of solar panel, and a genuinely unglamorous research question. Can the chips survive ten g of launch vibration, orbital radiation, and the thermal swing of a vacuum? The prototype can only run its accelerators in fifteen-minute bursts before they need to cool, which tells you how far this is from a data centre. Google has already irradiated Trillium TPUs on the ground at UC Davis; this is the first time the hardware gets shaken for real. The long version of the plan is constellations of up to 81 satellites flying in kilometre-wide formation at around 650km, talking over laser links, running on sunlight nobody has to buy — with two more satellites in 2027 just to test the lasers. Travis Beals, who runs Google’s Paradigms of Intelligence group, is refreshingly unromantic about next week: “This first launch is about seeing what works, identifying points of failure and applying those findings to future missions.” Every other compute story this year has been about permits, substations and water. This one is about whether a chip can take a beating.
Source: SiliconANGLE, September 24, 2026
An Agent Finally Beat NetHack, a Game Whose Players Consider Losing the Normal Outcome
On September 21, a GPT-6 Astra agent running through a custom terminal harness ascended in NetHack 3.6.7 — the full win, not a deep run — as a lawful dwarven Valkyrie, in 37,140 turns, for 1,766,446 points, on its third campaign. NetHack is permadeath, forty years old, and cruel in ways that are difficult to convey to people who have not personally starved to death holding a cursed tin opener; the community’s own wisdom is that most players never ascend. The repository is admirably careful about what it is claiming: “To our knowledge this is the first recorded LLM-agent ascension,” immediately qualified as “a historical-priority claim open to correction, not a benchmark result,” with a note that the non-LLM bot BotHack got there years ago. It also says outright that this is the harness, not a standalone autonomous player — the agent had web access, the wiki, and persistent memory across the run. Take it for what it is: not a benchmark, and not nothing. Thirty-seven thousand turns without a single unrecoverable mistake is a long time to hold a plan together.
Source: kenforthewin/nethack_astra on GitHub, September 21, 2026