The first model tested, Claude Opus 5.5, did draw up the experiment plan. It then refused to push the needle into the mouse's brain. The team had to switch to Claude Sonnet 5 to continue.
Founder Gabriel Ocana Santero described the procedure to WIRED. Once the vet had opened the skull, the AI chose the therapy, the brain region and the injection coordinates from a table of 19 options. It also set the injection volume and speed.
The team had put safeguards in place. A human had to approve the injection plan, an emergency button could cut the device's motors, and the AI was not allowed to touch the machine's internal software.
During the operation, the device's coordinate system failed. The AI saw that the injection position had been miscalculated, but the safeguards stopped it from correcting the code. The team therefore called in another AI, Opus 5, to repair it.
The experiment took a little over an hour, compared with about 45 minutes for a standard operation of this type. According to Gabriel Ocana Santero, the mouse is recovering well. The team is waiting for the therapy to take effect to measure its impact on the epileptic seizures.
Source
- 爱范儿 ifanr · ifanr.com