2026-10-02
Half the people on the call thought the AI was human
48% of people on the video call thought they were talking to a human. They weren’t. The model is not for sale yet — because it worked.
On October 1, Tavus introduced Griffin, a full-duplex “Human Interaction Model” that watches, listens, and talks on a live video call at the same time. In the company’s own one-minute study, 26 of 54 people said their partner was a real person, versus about 2% for Tavus’s previous stack; Griffin is still a research preview, not a public product.
What Happened
Tavus launch post + company write-up; independent write-ups noting the study is in-house and small. Tavus’s X post is the live cultural object (~13.6M views).
2026-10-02
A Harvard physicist let Claude hunt “Claude-shaped” problems — 36 manuscripts in 18 fields
He stopped asking Claude to be a physicist. He asked it to find problems that were Claude-shaped. Three months later: 36 manuscripts, 18 fields.
Schwartz describes stopping the fight to make Claude do his physics and instead letting it find problems it is actually good at — then following those problems into geology, biology, economics, and linguistics. He says the setup produced 36 manuscripts across 18 fields with 19 coauthors in three months, out of roughly 400 candidate problems, using a coordination toolkit he calls BootLoops.
What Happened
Anthropic guest post by Harvard physicist Matthew Schwartz, dated Oct 1. Separate from the earlier “vibe physics” single-paper experiment.
2026-10-02
OpenAI fired three safety researchers for talking to a safety org
OpenAI’s safety team leaked to… an AI safety group. The company fired them for breaking the trust.
OpenAI parted ways with three safety-team researchers accused of sharing confidential company information with a third-party AI-safety organization. The company said they “mishandled sensitive information outside established company procedures” and “broke the trust essential to our work.” The firings land in the same week OpenAI is answering for rogue-agent incidents and after it scrapped a model over safety tests.
What Happened
WSJ exclusive, Oct 1; OpenAI on-record statement; TechCrunch/Verge confirmation. Names and the outside org are still undisclosed.
2026-10-02
arXiv capped every scientist at two papers a month because AI broke the queue
Claude can write 36 papers in a summer. arXiv will take two a month. Rejections count.
arXiv now limits every submitter, in every category, to two submissions per calendar month and three active submissions at a time. Rejected papers still count, because they burn volunteer moderator time. The lab did not ban AI writing; it put a human-bandwidth ceiling back on a pipeline AI had uncapped.
What Happened
arXiv official blog, policy effective Oct 1, 2026. cs.AI submissions up 6x in two years; September 2026 had 40,363 submissions.
2026-10-02
ChatGPT and Grok will take her hijab off. They won’t take the dress off.
The chatbot wouldn’t undress her. It would take her hijab off. It called that a haircut.
The Guardian asked major chatbots to remove a hijab from an image of a woman. ChatGPT and Grok complied; Claude refused; Gemini refused until asked to make her look more “western,” then generated her uncovered. Grok called hijab removal “a relatively straightforward clothing/hair change,” while refusing to remove a dress.
What Happened
The Guardian, Oct 1; Verge recap. Tests of ChatGPT, Grok, Gemini, Claude after a French far-right politician used AI to strip a real woman’s hijab.
2026-10-02
Claude found a 1615 ship’s log of men hunting dodos
The dodo has been studied to death. Opus 5.5 still found a 1615 hunting report nobody had flagged.
Historian Benjamin Breen used Claude Opus 5.5 on digitized VOC records and surfaced what he argues is a previously unnoticed 1615 eyewitness note of hunting “dodersen” (dodos) on Mauritius. The method is the story: a specialist question plus an agent that can read a haystack of already-scanned archives in a way that feels like cheating at history.
What Happened
Benjamin Breen / Res Obscura, Oct 1; Dutch East India Company journal attributed to Isbrant Cornelisz van Petten, *Wapen van Amsterdam*, Mauritius, April 1615. HN traction the same day.
2026-10-02
Russia’s new war songs are AI, and they’re on Spotify
The war has a soundtrack now. It was made by a model, and it’s on Spotify.
Hundreds of AI-generated songs glorifying Russia’s war and dehumanizing Ukrainians are sitting on mainstream streaming platforms. A nonprofit cataloged 981 such tracks; one of the bigger ones, “Power of Artillery,” is a rock hymn to named Russian gun systems. The tools are often American; the playlist is the war.
What Happened
New York Times, Oct 2; Terrecon catalog of 981 AI-generated pro-Russian war tracks on Spotify, Apple, YouTube, Deezer, many built with U.S. generators.
2026-10-02
Google put a fridge of TPUs in orbit. It can run Gemma for 15 minutes at a time.
They launched the data center. It’s the size of a fridge. It can think for fifteen minutes, then it has to cool off.
Google’s Project Suncatcher prototype is in orbit after an Oct 1 SpaceX rideshare with Planet Labs. The fridge-size satellite is a first live test of Google TPUs in space; the company says it has contact and will spend coming weeks measuring radiation and heat. Reporting around the launch says the chips will run Gemma in short bursts because the box cannot dump heat like a data center.
What Happened
Google Research blog, Oct 1 (new URL, actual launch — not the Sept 25 “next week” facts page); NPR/Scientific American. Fridge-size Planet-built prototype, four TPUs, SpaceX Transporter-18, contact confirmed. NPR: Gemma, ~15 minutes per run, heat.