The Signal — September 20, 2026

Two weeks ago the frontier labs started talking publicly about slowing down. Today four consumers are in federal court arguing that the talking was the crime. Elsewhere, a Chinese research lab put a radiology model in Science and gave away the weights, and DeepMind found a way for agents to test strategies against a recording instead of paying to run them again.

Four consumers sued four labs for saying they should slow down

On September 18, four consumer plaintiffs filed a Sherman Act class action in the Northern District of California against Anthropic, OpenAI, SpaceXAI and Google. The case is Buist et al. v. Anthropic PBC et al., No. 3:26-cv-10693. The allegation is that when executives at those four companies publicly agreed the industry should pace the frontier, they were not making a safety argument. They were agreeing, as competitors, to restrain the products they sell.

Everything in the filing is an allegation. No court has found anything, no defendant has answered, and a complaint is the cheapest document in litigation to produce. What makes this one worth reading is that it points at a real gap. Antitrust law is built around competitors secretly agreeing to give customers less. It has very little to say about competitors openly agreeing to build something slower, in public, on stage, for stated safety reasons. The plaintiffs are betting those two things look identical from the customer's seat.

We covered the pacing statements themselves on September 13 and 14, when Dario Amodei made the proposal and the other lab heads endorsed it while Washington told them to calm down. This is what happens next. If the case survives a motion to dismiss, every future industry-wide safety commitment gets drafted by antitrust lawyers first, which is a strange and probably underrated constraint on how coordination in this industry can work at all.

Sources: POLITICO · Bloomberg Law · Los Angeles Times · The Hill

Alibaba published a CT-reading model in Science and released the weights the same day

Alibaba's DAMO Academy published DAMO RADAR in Science on September 18 and posted the code and weights to GitHub and Hugging Face at the same time. The model reads a contrast-enhanced abdominal CT scan and screens it in one pass for 146 conditions across 18 organs, including malignant tumors. On an internal cohort of roughly 39,000 real-world exams it reached a mean area under the curve of 0.913 across those findings, and the study reports it outperforming 23 of the 26 specialist radiologists it was measured against.

The reason a generalist model is possible here at all comes down to what it trained on. Most medical imaging AI learns from scans that physicians have annotated slice by slice, which is slow, expensive, and effectively caps how many conditions one model can cover. RADAR instead trains on the pairing that hospitals already generate in enormous volume: the CT volume and the radiologist's written report. That is vision-language contrastive learning applied to a problem where the labels were already sitting in the record, unread by anyone building models.

Three things to hold onto. The head-to-head comparison against radiologists comes from the authors' own study, which is normal for a Science paper and still not the same as independent replication. The validation cohorts are Chinese, so performance on other populations is an open question rather than a settled one. And releasing open weights is not a regulatory clearance; a hospital outside China that wants to put this near a patient still has an FDA submission and a local validation program ahead of it. What the release does change is the cost of finding out, which until now was the main barrier for everyone who is not a large hospital system.

Sources: South China Morning Post · TechTimes · Anadolu Agency

DeepMind's agents rehearse against a recording instead of rerunning the experiment

Researchers at Google and DeepMind released Dream-RSI on September 19, along with the code. It addresses a boring and expensive problem in self-improving agents: an agent searching a large space has to keep deciding which approaches to pursue, which to run in parallel, and which to abandon, and the only honest way to learn whether a different strategy would have worked better is to run the whole costly search again with that strategy.

Dream-RSI stores every attempt and its evaluation as the agent searches, building a record of the tree it explored. New strategies then get tested against that stored record rather than against the live model. The agent can ask what would have happened if it had chased a different branch first, or cut one earlier, and answer thousands of those questions without calling the model or the evaluator again. The researchers compare it to walking an unfamiliar neighborhood once, then planning a better route from the map instead of retracing every wrong turn.

The reported gain is up to 2.43x fewer generations on a VGG16 task and 1.79x on LayerNorm, which is a result on two specific benchmarks and not a general speedup. The more interesting constraint is what stays fixed. Nothing about the model itself moves: the weights, the evaluator and the tools all stay put, and only the search policy adapts, which means this is an agent getting better at looking rather than an agent getting better at thinking. It landed the same day Nathan Lambert published his argument for why he still does not buy true recursive self-improvement, and the pairing is accidental but clarifying: most of the visible progress in this area is search efficiency, which compounds slowly and predictably, not the fast-takeoff loop the phrase tends to summon.

Sources: The Decoder · arXiv · GitHub · Interconnects

On the Editor's Desk

Gemini Robotics 2 showed up in today's candidate pool with a September 19 timestamp, but the actual launch was July 30. That was an ingestion artifact, not news, so we left it out.

We also skipped the EFF's statement on Newsom's AI executive order, since we ran the executive order itself yesterday and a reaction to it does not add much. Anthropic's assessment of its four cybersecurity testing incidents is a genuinely important document, but it was published September 9 and what came out this week was commentary rather than new findings. And the EU AI Act's high-risk delay to December 2027 resurfaced again today, but the Council approved that back in June.

One we wanted and could not use: an arXiv paper arguing that current mitigations transform gender discrimination in GPT models rather than reducing it. It is a single preprint with nobody else having looked at it yet, and one unreplicated paper is not enough for us to put a claim like that in front of you. If it gets picked up or replicated, we will come back to it.