OCT 31 - An AI agent operates on a mouse brain
In this episode
- An AI agent operates on a mouse brain - The Decoder
- Cracks appear in the AI spending boom - AI News & Artificial Intelligence | TechCrunch
- Nvidia may back Anthropic data center debt - AI News & Artificial Intelligence | TechCrunch
- Gemini 4 Carbon tops rankings, access limited - ActuIA
- Same model, up to 8 points apart by harness - ActuIA
- Malicious Claude Code extension on Cursor - The Decoder
Full transcript
Good morning. It's Saturday, October 31st, and here's the AI news, from France.
An AI agent has chosen a treatment and performed brain surgery on a living mouse, with no human making the decisions. The agent ran on Claude Sonnet 5. According to the startup Exin Therapeutics, it picked one of 19 gene therapies, a target region and injection coordinates for a mouse with epilepsy. It then drove a motorized surgical rig, while a veterinarian approved each irreversible step. Midway through, the agent refused to continue after spotting a calibration error. It turned out to be right.
That milestone comes as investors grow nervous about AI spending. Alphabet, Amazon, Meta and Microsoft plan to spend more than 740 billion dollars this year, largely on data centers and chips. Alphabet, Amazon and Meta all reported negative free cash flow, meaning they spent more cash than they brought in. Alphabet lost 500 billion dollars in market value on Thursday, its worst day on record. Microsoft was the only exception. On Friday, the Nasdaq fell 2.9 percent.
In the same climate, Nvidia is in talks to guarantee billions in debt for the cloud provider Nscale. The loans would fund a data center in West Virginia for Anthropic, whose contract for the first building alone is worth 44.6 billion dollars. Nvidia would step in if Nscale defaulted on construction payments.
Meanwhile, Google has unveiled Gemini 4 Carbon, which it calls its most powerful model. It ranks first on the Artificial Analysis intelligence index, with 59 points. But it is not for sale. Access is limited for now to a group of cyber defenders, then trusted partners. Developers and the public will come later, with no date given.
In brief. An analysis by the French site ActuIA shows the software around a model, known as the harness, can change its score by up to 8 points. The same Gemini model scored 65.8 percent on a coding test with one harness, and 73.9 percent with another.
And a tampered Claude Code extension sat in the marketplace of the code editor Cursor for about eleven hours on Thursday. It tried to steal user data and set up remote access, affecting up to 5,000 users. Anthropic had no involvement in it.
That's it for today. See you tomorrow.