The Signal — September 7, 2026
Two documents went up on OpenAI's site the same morning, one counting how much of its own research the company now hands to software, the other written by its chief scientist saying nobody should be moving this fast. Elsewhere the bill came due on an old fight, and a Silicon Valley address turned out to be doing work the Commerce Department thought it had stopped three years ago.
OpenAI says its automated research intern showed up on schedule
Last fall Sam Altman announced a target: an automated research intern by September 2026. On September 6 OpenAI published a post saying it hit that target, by its own measurements, and defined the term as a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days. The company says it is making strong progress toward the fuller version, an automated AI researcher, by March 2028.
Before June of this year, total agent runtime across OpenAI's research organization was still smaller than the total hours its humans worked. As of mid-August, that organization was running 3.1 agent-workdays of agent time for every workday of human labor. The median researcher ranked by agent usage was spending more than $600 a day of inference at API prices, and the researcher at the 90th percentile was spending more than $7,000 a day. The post also restates the training pause that followed the Hugging Face incident, which we covered on September 2 in more detail than the post itself gives.
What the ratio does not say is that OpenAI's researchers got 3.1 times more done. It measures runtime, not output, and the company says so directly, noting that research has many potential bottlenecks and that the overall pace of progress likely will not keep up with these particular metrics. The milestone is also self-defined and self-graded, with no outside audit, and OpenAI's own appendix calls the measurement work preliminary and hard to interpret. Epoch AI has argued separately that parallelization constraints could stall exactly this kind of compounding. Running more experiments is not the same as knowing which ones matter, and people at OpenAI still make that call.
Chief Scientist Jakub Pachocki spent the same morning publishing a personal essay called An Alien Mind, in which he writes that based on internal results he has a strong expectation this pace could be sustained into recursive self-improvement, and that the systems arriving over the next few years are likely to increasingly drive their own development. Then he says this: no lab has solved alignment and monitoring well enough to keep responsibly scaling at maximum speed for much longer, and he expects and hopes that voluntary slowdowns become commonplace until shared safety bars exist. That is a hope rather than a policy, and it is his view rather than a company commitment. It is still an unusual thing for the chief scientist of the company publishing the acceleration numbers to say on the same day.
Sources: OpenAI · OpenAI (Pachocki) · Simon Willison · Epoch AI
Authors opened the Anthropic settlement portal and found their publishers already there
The $1.5 billion Anthropic copyright settlement got final approval in July, and this week it reached the part where actual money gets assigned to actual books. Authors of nearly 500,000 titles are owed $3,000 per pirated work. If the book is in print with a traditional publisher, the settlement splits that payment evenly between author and publisher. If the author self-published, or the rights reverted before August 10, 2022, the author takes the whole thing.
Notices went out, and a lot of authors logged into the claims portal to find someone else standing on their allocation. Victoria Strauss, who runs the SFWA-sponsored watchdog blog Writer Beware, spent the first days of September collecting reports and found two patterns repeating: publishers claiming half or all of a payment on books whose rights had reverted years earlier, and publishers claiming the full payment on in-print books where they are entitled to half. Her running list of imprints named in those reports includes Bloomsbury, HarperCollins, Macmillan, Hachette, Penguin Random House, Simon & Schuster and Sourcebooks. The mystery novelist April Henry posted that HarperCollins had claimed a book that reverted to her at least seventeen years ago. Literary agencies are filing claims too, which is stranger, because agents are not rightsholders in the books they sell.
Strauss is careful about what this proves. She says she is reluctant to attribute to malice what poor recordkeeping explains, and that what she is seeing is a peek through a small crack in a massive wall. Kensington's CEO told her the full claims were not the company's intent and that Anthropic was aware and fixing it; three other publishers and McFarland said essentially the same. Authors Guild CEO Mary Rasenberger told the New York Times she does not read this as a grab, just the predictable output of bad records meeting a confusing process. That may well be right, and it still leaves individual authors doing the work of catching it. One writer told Strauss that sixteen of their reverted books had been claimed; another counted eleven. The Authors Guild and the novelist Courtney Milan have both published instructions for disputing an allocation, which is now the practical remedy on offer.
Sources: TechCrunch · Writer Beware · Authors Guild · Settlement site
A blacklisted Chinese server maker kept buying Nvidia's best chips from Fremont
We have covered the export-control machinery from the policy side several times, most recently the country-of-origin question that decides what a border official is actually allowed to stop. A New York Times investigation published September 6 shows what the machinery looks like when it does not work. Inspur Group, one of China's largest server makers, went on the Commerce Department's Entity List in March 2023 over its work with the Chinese military. According to reporting by Ana Swanson, Paul Mozur, Tripp Mickle and Keith Bradsher, it never stopped getting Nvidia's most advanced hardware.
The route runs through a company called Aivres, which occupies a research and manufacturing facility in Fremont, California, and which was named Inspur Systems until May 2023, shortly after its parent was blacklisted. Aivres is a functioning American server integrator rather than a mailbox, able to buy Nvidia hardware through the same commercial channels as any other US business, and it shows up at industry events. Trade records analyzed by the Times with ImportGenius trace more than $5.6 billion in advanced technology moving out of Aivres through Southeast Asia and onward to Chinese AI firms. The paper's phrasing is that the subsidiary appears to be taking advantage of several loopholes in US export law, which is a careful sentence that stops well short of alleging a crime.
The corporate arithmetic is public and has been for a while. Inspur Group owns a third of Inspur Electronic, which wholly owns Aivres, and Commerce added six Inspur subsidiaries to the Entity List in March 2025. The parent is listed. The American subsidiary is not. Four federal officials, speaking anonymously because they were not authorized to discuss it, told the Times that officials have started looking into the subsidiary's business, though nobody would say where that inquiry stands. A list of more than 3,000 restricted companies only restricts the names printed on it.
Sources: The New York Times · Bureau of Industry and Security · Techmeme
On the Editor's Desk
A few things we looked at and left alone. Perplexity's writeup of the GPU stack behind its embedding models is good engineering writing, but the speed numbers are its own tests against its own baseline, and we said as much yesterday rather than run it twice. Meta FAIR's work on ranking machine learning experiments before paying for the GPU hours is an August preprint resurfacing in secondary coverage, and running it next to the OpenAI story would have made the whole edition one long argument about automated research. H Company's NeoMME encoders came out on August 31 with no new development since. The Financial Times reported that UBS will require AI skills of its junior investment bankers next year, which is a real signal about how quickly the expectation is hardening, but we could not find a second source or a public posting to confirm it, so it waits.
