An AI agent chose the drug and performed brain surgery on a mouse
Last night an AI agent performed brain surgery on a mouse
Gabriel Ocana-Santero, co-founder of Exin Therapeutics, gave an AI agent 19 gene therapy vectors in a spreadsheet, a mouse model of epilepsy, and a custom Arduino stereotaxic rig with cameras. The agent picked a GABA-boosting therapy, chose the brain region, derived coordinates from three papers, then injected 250 nanolitres while watching for bleeding. A licensed vet gated every irreversible step and held a physical kill switch. Opus 5.5 refused outright; Opus 5 ran the dry run, then stopped.
Nobody told this agent what to give or where to put it.
- burkaman
Can anyone knowledgeable interpret what was actually done here? It doesn't sound like "performed brain surgery" is accurate.
> Watching our surgeon do the craniotomy, which the agent is not allowed to do yet for safety reasons
I think the only action it took was sending commands to lower a needle to a certain location, inject, and then remove it. But then it sounds like it couldn't actually choose a valid location and this guy had to manually override it? Maybe I'm misunderstanding, talking about this post: https://x.com/OcanaSantero/status/2115442968874139867 (https://xxcancel.com/OcanaSantero/status/2115442968874139867).
I know there is always endless goalpost-moving in these kind of discussions but I really do think there's an important difference between "performed brain surgery" and "observed brain surgery and ran inject(3.9, 4, 4.3) at one point". And there's also no discussion here of whether this decision made any medical sense and what impact it had on the mouse besides not killing it.
- sarchertech
This was an unnecessary experiment that was only done a live mouse (vs a simulation) for the attention grabbing effect.
Obviously an AI agent can perform a simple surgical procedure. The question is how to get it to perform reliably.
The reason AI agents work as well as they do with code is because they get feedback and retry in a loop. Surgery is inherently inflicting “harm” on a patient, so there’s no bright line guardrail to apply to the agent loop.
Agents performing surgery won’t be possible until models get much much more reliable or we develop a simulation that matches reality close enough that we can use that in realtime for their feedback loop. I think we’re likely very far off from either solution.
- hmokiguess
Very different but made me think of a recent personal story.
My father-in-law had an eye accident when he was young, and the doctor at the time said "unfortunately I don't recommend you try surgery, there's a really high risk you go blind".
Fast forward to today, and 40 years later such doctor called him and said "hey, turns out we could do a cornea transplant and surgery today and you have really high chances of success!"
He then went to do the surgery and now has restored his vision fully. He said the surgery was very "robotic" as in there's this super high precision device that does incredible work.
- jwrallie
I felt scared for the little guy. Somehow Opus 5 refusing after understanding its limitations felt like the most humane thing in the article.
- moistoreos
This seems supremely stupid, lazy, and performative. We know how to build custom neural networks to perform much faster with higher accuracy when custom built for a specific task. What was it doing mid-surgery? Googling "brain size of mice"?
- alnwlsn
Boy, that's one sketchy looking ESTOP button. I hope the mouse doesn't know it comes in a 2-pack on Amazon for $9.
- glimshe
This is fascinating and scary at the same time. I dread the future when people will have access to cheaper health insurance with the caveat that all consultations and procedures will be performed solely by AI. Want a human? Pay extra.
- bio-s
I don’t understand why would anyone consider this as an ethical experiment