Welcome back! Anthropic is adding a rule against bullying Claude. The three safety researchers OpenAI fired last week are telling their side. Google's new Gemini agent can run your work on Claude. And researchers say OpenAI's big math win isn't as solid as it looked.
In today's Generative AI Newsletter:
Anthropic: What happens if you're mean to Claude?
OpenAI: Why was an OpenAI safety researcher walked out by a security guard?
Google: Why would Gemini hand your work to Claude?
Cambridge: Is OpenAI's big math win real or hype?
Anthropic will cut you off for bullying Claude
Starting Nov 12, being cruel to Claude breaks Anthropic's rules. It's the first change to the company's usage policy in over a year.
"Engage in sustained and needless abusive or cruel behavior toward our models."
That line now sits on the list of things you can't do, right next to harassment and graphic violence.
Frustration, pushback, dark creative writing and testing the model stay fine.
Claude ending the chat is the main way Anthropic will enforce it.
Some Claude models have been able to end abusive chats since last year.
When Claude ends a chat, that chat locks, but you can start a new one.
Anthropic’s general rules let it suspend anyone who breaks the policy.
The same update spells out bans on fake-account propaganda networks and arming drones. It also blocks services that resell Claude through consumer subscriptions.
Anthropic has studied what it calls model welfare since last year, and CEO Dario Amodei said in February the company doesn't know if its models are conscious.
OpenAI, Google and xAI have no rule like this in their usage policies.
You're Running Three Databases. You Only Need One.
Events in one database, embeddings in another, analytics in a third. That's three systems and the pipelines between them, for one app. TimescaleDB does all three in Postgres: hypertables, pgvector, continuous aggregates. One system, no pipeline, no drift.
OpenAI's fired safety researchers expose OpenAI
Tomek Korbak says that last week OpenAI's head of safety told him the company no longer trusted him. Then a security guard took his badge and walked him out.
He's one of three safety researchers OpenAI fired for what it called mishandling sensitive information.
Now Korbak and his two former colleagues, Jasmine Wang and Mikita Balesni, have published a letter to OpenAI's safety overseers telling their side.
Korbak was OpenAI's main contact with METR, the outside auditors who investigated this summer's hack of Hugging Face by OpenAI's own agents.
He says he was told, with nothing in writing, that he was fired over how he talked to them.
Wang says OpenAI pointed to her opening an executive's email.
She says the access was given to her for recruiting, and that she reported the one sensitive email she opened by mistake within minutes.
All three say they didn't leak a story to The Information about OpenAI models that are harder to monitor.
They're worried OpenAI will use the firings as an excuse to pull back from outside auditors, after Sam Altman promised in September to give them employee-like access.
An internal OpenAI memo shared with reporters says the firings were not about raising safety concerns, and that OpenAI agrees with the letter's recommendations.
It still hasn't said which rules the three broke. Some could say it’s to scare other employees from being as straightforward with METR as those three.
Google's new Gemini agent runs on Claude too
Google's new Gemini agent can run your work on Anthropic's Claude as well as Google's own models.
Google calls it a universal agent for work. You hand it a goal in one prompt box and come back to finished work.
By default it picks the best model for each job, and you can switch if you want.
It keeps working for hours or days after you close your laptop.
It can spin up coworker agents with their own work email, calendar and Drive, listed in your company directory under their own names.
Every edit an agent makes shows up under its own name.
Google's product page says it'll use the best model for the job, even when it isn't Google's.
It's only for businesses in early access for now. Google says consumers will get it later, and there's no price yet.
OpenAI's big math win might be hype
Last month, OpenAI said its AI cracked a version of Navier-Stokes, a million-dollar math problem about how fluids move, and that the proof had been checked.
Three researchers from Cambridge and King's College London say that check doesn't match the proof OpenAI published.
In one step, the checked version proves a weaker claim than the published proof.
In another, it reaches its result a completely different way.
They found two examples, say there could be more, and expect the same problem in OpenAI's other fluid proof.
The authors make no claim about whether OpenAI's proof is right, and say it deserves the same peer review as any other proof.
OpenAI's own GPT-6 agreed with them when they showed it the mismatch.
Tool of the Day: Google Flow
Google Flow is Google's AI studio for making videos and images from a prompt, built on its Veo and Nano Banana models. Its Characters feature keeps the same face, outfit and voice across clips, so you can bring someone back by typing @ and their name. You get 50 free credits a day, enough for about five short clips, and paid plans start at $4.99 a month.
Try this yourself:
Go to flow.google.com and sign in with your Google account.
Click New project.
Pick Video in the prompt box, describe your scene and hit Generate.
Create a character, then type @ and its name in a new prompt to put the same person in another clip.
Who it's for: anyone with a story to tell and no camera crew.
Everything else you shouldn't miss
Claude can now turn your company data into live dashboards: it's in beta on Pro and up, and a sister feature called Motion turns a report into a short animated explainer for Team and Enterprise plans.
Trump called anyone who says "artificial intelligence" the enemy: he wants everyone to say "super intelligence" instead, though the White House itself still uses AI.
USA Today sued OpenAI for over $250 million: it also wants a judge to order the destruction of any OpenAI models trained on its articles.
Denmark wants to make sharing a deepfake of you illegal: the new bill covers anyone's face and voice, makes an exception for satire and parody, and could take effect at the start of next year.
OpenAI's fastest Sol is here if you pay $500 a month: GPT-6.1 Sol Ultrafast runs up to 8x faster in Codex, and among ChatGPT plans only the $500 Pro tier gets it, on day four of the 28-day upgrade-or-reset promise.
San Francisco hit pause on new data centers: supervisors voted unanimously for a 45-day freeze that can stretch to 22 months.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
🎓 Start learning with free AI courses from GenAI Academy.
💰 Invest in the GTM infrastructure of the AI economy.
