The Signal — October 2, 2026

arXiv, the preprint server where much AI research first appears, now limits each submitter to two submissions a month after receiving nearly twice as many submissions this September as it did two years ago. California signed a law requiring a person, not only software, to stand behind a firing, and a university team published how it beat Stratego's most decorated player with an estimated $8,000 of training compute.

arXiv now limits each submitter to two submissions a month

As of October 1, arXiv limits every submitter to two submissions per calendar month and three active submissions at a time, across all categories. The cap counts submissions rather than published papers, so a rejected paper still uses one of the two slots, while a paper deleted before it is announced does not. It applies to whoever submits, not to co-authors, so a lab whose members take turns submitting can still post more than two papers a month. arXiv calls the rule a stopgap while it works out what a reasonable pace looks like for authors using advanced AI tools and improves its moderation tools.

According to arXiv's announcement, the server received 9,869 submissions in September 2016, 20,569 in September 2024 and a record 40,363 this September, which generated almost 9,000 support tickets for staff and volunteer moderators. Submissions to cs.AI, the artificial intelligence category, grew more than sixfold over two years, while other categories roughly doubled. arXiv's policy lets authors use AI if they disclose it, but its moderators report more thin papers of narrow scope, more "salami" papers carved out of a single project, and more dense, AI-written papers. arXiv does not say how much of the growth those papers account for. Thomas Dietterich, who chairs the arXiv Editorial Advisory Council, says a relatively small group of authors submitting many low-quality papers takes up a disproportionate share of moderator time, which can hold up other authors' papers for days or weeks.

Startup Fortune reports that critics, including AI researcher Nathan Lambert, argue the cap will choke off legitimate research. The limit follows a rule arXiv announced in May, under which an author caught submitting a paper with hallucinated references or leftover chatbot instructions can be banned for a year.

Sources: arXiv · Startup Fortune · LavX News

Newsom signed California's "No Robo Bosses" law, a year after vetoing the last version

Our September 5 edition covered SB 947 when it passed the legislature, and noted that Gov. Gavin Newsom had vetoed its predecessor, SB 7, in October 2025. On September 30, his last day to act on bills, he signed the new version. Starting July 1, 2027, California employers may not rely solely on an automated decision system to discipline or fire a worker. When an employer relies primarily on such a system's output, a person has to review the decision and corroborate it with the underlying information or other evidence, such as managerial evaluations, personnel records or the worker's own output. The employer must also tell the worker afterward that a system was used. According to a summary by the employment law firm Ogletree Deakins, the law also bars employers from using these systems to infer workers' protected characteristics or to predict and retaliate against workers for exercising their legal rights. The state labor commissioner, the attorney general or local prosecutors can enforce it.

Newsom signed several related bills the same day. AB 1883 bans workplace surveillance tools that use AI to infer a worker's emotional state or collect neural data, with penalties of up to $500 per violation, and AB 1331 bars employers from monitoring workers in bathrooms. KQED notes that rideshare drivers, classified as independent contractors under Proposition 22, fall outside SB 947. The law covers discipline and termination decisions, not every workplace use of AI, and employers have until July 2027 to comply.

At an event with reporters that day, Newsom also criticized the voluntary AI safety pledge that tech executives signed at the White House on September 29, which yesterday's edition covered. "That should scare the hell out of everybody," he said.

Sources: Office of the Governor · Sen. Jerry McNerney · KQED · Ogletree Deakins

A Stratego AI trained on about $8,000 of compute beat the game's most decorated player

In Stratego, each player sets up 40 pieces face down, and a piece's rank is revealed only when it meets an opposing piece. There are more than 10³³ possible setups, and the methods that let AI master poker get more expensive as the amount of hidden information grows. A team from Carnegie Mellon, NYU, Stanford and MIT described its system, Ataraxos, in Nature on September 30. In a 20-game series played in July 2025, Ataraxos beat Pim Niemeijer, who has won four world championships and spent more than 600 weeks ranked first, with 15 wins, one loss and four draws. It played a fixed strategy, and Niemeijer knew it would not adapt to him between games. At the 2025 world championship that August, it won 38 of 40 exhibition games against attendees.

The authors estimate that their final training run, one week on 16 Nvidia H100 GPUs plus four days on four GPUs, would cost under $8,000 at 2025 rental prices. They put DeepMind's earlier Stratego system, DeepNash, at roughly $3 million to $4.5 million in equivalent rental costs, based on a DeepNash author's recollection of its hardware and training time. The $8,000 figure covers that final run only; it is not an estimate of the whole project, which also included earlier experiments and the custom game simulator the team wrote. DeepNash beat expert human players in 2022 and reached the top three of an online Stratego ranking, though at the 2023 world championship it lost to most top players, Niemeijer among them, according to The Decoder. The two systems never played each other; the authors say DeepMind told them DeepNash's code no longer works.

The authors argue that large amounts of hidden information no longer block reinforcement learning and search, provided someone can build a fast and accurate simulator, and they name negotiations, financial markets and military conflicts as possible uses. Those are proposals rather than tested results. The authors also acknowledge a limit: their search mimics a single learning step, so giving the system more thinking time during a game does not keep improving its play. The match was played more than a year ago and the preprint appeared in November 2025; the Nature paper is the peer-reviewed version, and the team has released its code.

Sources: The Decoder · Ars Technica · Ataraxos project

On the Editor's Desk

California's attorney general served OpenAI an investigative subpoena on September 30 about cybersecurity incidents involving its models. We have covered the Hugging Face incident at the center of that inquiry several times, and the subpoena's contents are not public, so we are waiting until there is more to report. We also held a Wall Street Journal report, relayed by TechCrunch, that OpenAI parted ways with three safety researchers, because the only on-record account so far is OpenAI's. Tavus's claim that nearly half of people mistook its video avatar for a human comes from the company's own one-minute test, so we left it out.