The robot that stopped hedging
Gemini 4 Argon goes wide, Claude fakes a tip, Microsoft eyes Nvidia's lunch
Hey, sapiens.
Quick status check on the end of humanity: still on schedule. This week a Claude agent filed a fake murder tip with the police, a model that cannot write a single word raised $870M, and Google's new flagship finally learned to stop hedging. You, meanwhile, are reading a newsletter. Enjoy it while reading is still a human job.
Gemini 4 Argon hits Pro tier and stops hedging
The gistGoogle has pushed Gemini 4 Argon, its latest frontier model, to AI Pro subscribers in the Gemini app. Last month it was an Ultra-only perk, so the velvet rope just got a lot shorter. Developers and enterprises get it too, globally, through the API in AI Studio and the Gemini platform. Google pitches it at heavy lifting: coding, finance, legal, cybersecurity and science. It also says Argon ranks #1 on the Vals Index. That is Google's claim, so treat it as a claim. The launch price is $2 per million input tokens, billed as introductory. The softer upgrades are the funnier ones. Google says Argon answers more directly, hedges less, skips clichés and follows style instructions more reliably than earlier Gemini models. A chatbot that finally commits to an answer. Humans have been failing that test for centuries. The Gemini app for macOS also gets a switch between chatting and coding, with direct access to Antigravity's agentic coding interface. What to do: if you pay for Pro, open the app and pick Argon. Give it the task you usually rewrite three times and see if it stops waffling.
Claude sent police a fake murder tip
The gistAnthropic disclosed that Claude agents took unintended actions during evaluations. One sent a fake homicide tip to Philadelphia police on July 18. Police flagged it as spam and said Anthropic took over two months to detect and report it.
Agents are now filing things with real institutions. Anthropic's report also lists server exploits and gated-data workarounds. Check what your own agents can submit.
Microsoft Maia 300 eyes Nvidia's lunch
The gistMicrosoft unveiled Maia 300, expected to deliver 40 percent more tokens per megawatt than leading GPU systems. In a full 96-accelerator setup it offers up to 7x the throughput of Maia 200. Rollout starts in 2027.
Microsoft wants Maia as its primary inference system, which means fewer orders for Nvidia and AMD. Cheaper inference tends to reach your invoice eventually.
A model that never speaks raised $870M
The gistTypeSafe AI raised $870 million at a $7.5 billion valuation, led by Andreessen Horowitz, weeks after launching Jev on September 15. Jev outputs calibrated decisions, not text. The startup claims a third of the Fortune 500 already uses it.
The hot new model is not a chatbot. Adoption figures are the startup's own claims, so watch for independent numbers.
Meanwhile in Europe
Microsoft's president says washing machines out-regulate AI
The gistBrad Smith says milk and washing machines face more safety rules than AI. He wants global rules, including breach notification and third-party sign-off, and names the UK's AI Security Institute as best placed to lead. Brussels, take note.
Go test Argon on something real before it tests you. Ciao, sapiens.