Blog
Writing about AI — how it works, where it helps, and what we learn building with it.
Latest posts
OpenAI Has Shipped Text Watermarking. Only the EU Gets It by Default, and Nobody Outside an Approved List Gets the DetectorOpenAI's textGrain watermark became automatic for EU ChatGPT and Codex users on October 5 and stayed opt-in everywhere else. Its published detection rates show a signal that survives reading and not editing — and the detector is available by application only.Artificial intelligence
Trump's AI Force Now Exists. It Is Chaired by the Director of National Intelligence and Told to Prevent OverregulationTrump named the Super Intelligence Force on Sunday: DNI Jay Clayton as chair, the FTC, Pentagon and OPM as vice chairs, and 120 days to report. The charter asks it to plan for AI-enabled threats while preventing overregulation — and to review the breach reporting the White House accord left out five days earlier.Artificial intelligence
OpenAI Fired Three Safety Researchers for Sharing Information Outside the Company. One Was Its Own Contact for the Outside InvestigationOpenAI confirmed it dismissed three safety staff for mishandling sensitive information, without naming them, the outside group, or the material. Reporting identifies the recipient as METR and one of the three as OpenAI's own technical contact for METR's Hugging Face review.Artificial intelligence
OpenAI Says Its Hidden Reasoning Was Extracted by Asking the Model to Transcribe It. The Encryption Held; the Model Did NotOpenAI's 30 September report attributes a coordinated reasoning-extraction campaign to individuals associated with Moonshot AI. The numbers are small next to Anthropic's September accusations; the method is not — the operators replayed encrypted reasoning and asked a model to transcribe it, which is the one failure a stronger cipher does not fix.Artificial intelligence
Tencent Is Reported to Have Rented 100,000 Chips It Cannot Buy. Export Controls Regulate Where Compute Sits, Not Who Uses ItThe Financial Times reported on September 30 that Tencent has leased about 100,000 advanced AI chips from Oracle for roughly $7 billion over five years, hosted in Southeast Asia. Tencent cannot buy those chips; renting them offshore is legal, and the House bill written to close that gap has been sitting in the Senate since January.Artificial intelligence
Google's First Frontier Model Since Gemini 3 Went to Cyber Defenders Before It Went to Customers. The Order Is the AnnouncementGoogle says Gemini 4 Argon leads or ties on 13 of 18 benchmarks. Nobody outside a vetted programme can check, because the model went first to cyber defenders and a US government review process and only later to the people paying for tokens.Artificial intelligence
The FTC's First Rogue-Agent Probe Names METR Alongside OpenAI and Anthropic. The Auditor Was Supposed to Be the Answer, Not a SubjectThe FTC confirmed on Wednesday that it is investigating OpenAI, Anthropic and other AI developers over the consumer risks of autonomous agents, using the 1914 FTC Act rather than any new AI law. The third name on the list is METR — the nonprofit evaluator the industry offered as its own check.Artificial intelligence
DeepSeek Open-Sourced Six Libraries for Huawei's Ascend Chips. The Software Layer Is the Part Export Controls Never TouchedDeepSeek published a TileLang backend for Huawei's Ascend 950 and five kernel and communication libraries on September 30, developed with Huawei and merged upstream next to the CUDA backend. Export policy restricted the chips; it never restricted CUDA, and CUDA was always the harder thing to replace.Artificial intelligence
OpenAI Announced More Than Twenty Things and Only One Was a Model. The Rest Is an Attempt to Become the Place Software LivesOpenAI's DevDay shipped a word processor, a slide editor, a shared drive, an identity provider, a plugin UI surface and an enterprise procurement channel. What it did not ship is a billing system — and an app store without a cash register is a subsidy, not a business.Artificial intelligence