The AI slowdown pact draws an antitrust suit

Four people who pay for ChatGPT, Claude, Grok or Gemini sued the four companies on Friday in federal court in Northern California, seeking to represent every paying subscriber in the country. The complaint says the rivals agreed to slow AI development together when Sam Altman, Elon Musk and Demis Hassabis endorsed Dario Amodei's slowdown proposal on September 12, and traces the coordination back to a July statement signed by senior lab employees. Each company may slow down on its own, the plaintiffs argue, but competitors may not agree to. Amodei had asked Washington for a narrow antitrust waiver for exactly these talks, and OpenAI said it did not need one. None of the companies had responded by Saturday.

Nvidia chief executive Jensen Huang told CBS News there is a zero percent chance AI ends the world by 2030, called warnings of that kind irresponsible, and said calls from Dario Amodei and Sam Altman to slow development are not grounded in science. He argued that existing product liability and cybersecurity law is regulation enough, and told communities angry about data centers, "we're sorry we didn't come talk to you sooner." A CBS News and YouGov poll of 2,051 adults, taken the same week, found 43 percent want AI companies to slow down, 12 percent want them to stop, 37 percent favor the current pace and 7 percent want it faster. Sixty-one percent said the companies will not make sure AI is used appropriately.

Alibaba's Qwen team published Qwen-Image-2.1 on Hugging Face, a single model for generating and editing images with a 7-billion-parameter image generator, native transparent backgrounds, and edits guided by up to 10 reference pictures. The weights come under a new Qwen Research License that allows research and evaluation only. Any commercial use needs a separate license from Alibaba, and disputes go to courts in Hangzhou. Earlier Qwen-Image checkpoints were Apache 2.0, and the Qwen-Image-3.0 flagship announced in July shipped no weights at all, so Alibaba's image models now carry three different sets of terms.

ConvAI Innovations released Laya on Hugging Face under Apache 2.0, three small checkpoints of 322 million to 421 million parameters that, like TypeSafe's Jev, answer typed questions about an email, ticket or document with a probability attached rather than writing text. Its model card reports about 33 milliseconds per decision on a single T4 GPU, against the 236 to 276 milliseconds that outside testers measured for Jev, and higher scores on three of four classification tests. It is candid about the limits: the Jev figures come from other people's tests, and Jev wins easily when there are dozens of labels, 87 percent against 43 percent on a banking benchmark.

In a guest post on Terence Tao's blog, mathematician Po-Shen Loh answers critics of this month's open letters from mathematicians against AI labs. The Leiden Declaration alone has more than 4,000 signatures. He starts from one axiom, that humans should help humanity flourish, and one observation: no far more capable species has ever handed control of its future to a less capable one. If every field insists that accountable people supervise the decisions AI makes, he argues, that oversight creates more work than there are people to do it, and the shortage, not regulation, is what eventually slows AI down.