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.
I don't know what harnesses they run but when I daily tell Claude or others to do X and they indeed do parts of X and then also Y and break A on the way... I'm so tired of asking "but, really?" that I'm not even bothering anymore. The AI folks live in a different world than I live, and whether it's a real world or only in their minds (or only in their articles) I definitely cannot tell anymore. I simply gave up.
Some variation of “you’re just using it wrong” is usually the refrain I hear when I talk about things LLM’s are bad at. Anything is possible if you divine the right words to input and fix your harness and set pick the perfect model and…
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.
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.
This is already happening though, to a point; recently, local GPs have been subsumed into larger health care provider companies, and they have set up systems so that before you can make an appointment, you need to go through a wizard and / or chat interface. Likewise, people are encouraged to go to a website and do some self-diagnosis before calling a doctor - things like "if you have a fever, how high is it and for how long have you had it?"
It's indirectly related to health insurance cost, as we have a GP shortage and they (I presume?) realized a lot of people are making appointments and wasting a GP's time for trivialties like a cough or cold.
I'm pretty sure they already have, or are experimenting with having people go through an AI chatbot first. This is already very much a thing with any kind of customer support. It's not a new development though, the preceding generation was conversational interfaces.
I imagine human consultation could be more expensive, at least in the US, due to liability of human error. Especially if/when AI capabilities in medical procedures becomes clearly superior.
Humans are irrational. We also value "feeling" safe. Some tricks sell the illusion for those people who can't trust 100% some technologies.
Of course I trust engineering, I know planes are safe: but when there is turbulance and I see a flight attendant just casually walking and not even reacting to it, I feel better.
Humans are still humans.
Of course, culture changes. Colective ideas and institutions adapt. But extremely slowly compared to advances in tech.
I feel like we are in the grab headlines by any means necessary part of the AI Hype cycle, maybe a bearish indicator.. either way this just felt wrong. Like what the heck kind of wrong.
This truly feels like a meme rather than a genuine study or something... If I hook up claude to my car it will probably will be able to drive it, same way it can program a simulated car in a game engine, or even Jev could do it...
This has be some sort of a weird marketing stunt for so called x-bros that I can't understand.
It’s a mouse. If you think feeding one to a snake is ethical, then it’s a little strange this would be unethical. At least the human has the mouse’s wellbeing at heart.
That’s not quite equivalent. For one, a responsible owner is not feeding live mice to a snake. There’s at least an opportunity to ethically euthanize the mice prior to the act.
> At least the human has the mouse’s wellbeing at heart.
Not saying that this resolves the broader issue, but I'm not aware of any vegetarian snakes. It seems possible to differentiate this from animal experimentation.
How much issue do you take with animal testing in the rest of medicine? We already, more or less, torture an unfathomable amount of animals to get where we are.
That ship sailed a long time ago. Science decided that killing mice for experiments (regardless of the experiment) is fine. In fact, a large amount of the mice killed are just redundant (not needed for a particular experiment)
I would feel more comfortable if the agent wrote the software to do the surgery rather than did it itself, so that it can be vetted before deployment. And if a complication arose, to stop and defer to the surgeon (would that work?)
If only we had accurate models they could train on.
> 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.
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.
It's indirectly related to health insurance cost, as we have a GP shortage and they (I presume?) realized a lot of people are making appointments and wasting a GP's time for trivialties like a cough or cold.
I'm pretty sure they already have, or are experimenting with having people go through an AI chatbot first. This is already very much a thing with any kind of customer support. It's not a new development though, the preceding generation was conversational interfaces.
Need an ambulance? Let me escalate to the department that will escalate you to a department that will dispatch the ambulance in 5-10 business days.
Of course I trust engineering, I know planes are safe: but when there is turbulance and I see a flight attendant just casually walking and not even reacting to it, I feel better.
Humans are still humans.
Of course, culture changes. Colective ideas and institutions adapt. But extremely slowly compared to advances in tech.
This is an unfortunately true and insightful comment.
I noticed that the first tweet made no mention if the mouse survived. Seemed a little...sus.
Without more context, I’m not sure if that’s better or worse
Last night an AI agent performed brain surgery on a mouse, yeah
'Cause I was sittin' there bored to death
And in just 18 hours in front of my laptop
It picked the drug
It picked the brain region
It drove the stereotactic arms into the brain and injected 250 nanoliters while watching the entry site through the rig cameras for bleeding, girl
https://www.amazon.com/dp/B09JP4JKY6
This truly feels like a meme rather than a genuine study or something... If I hook up claude to my car it will probably will be able to drive it, same way it can program a simulated car in a game engine, or even Jev could do it...
This has be some sort of a weird marketing stunt for so called x-bros that I can't understand.
Will we just throw LLMs at most problems in the future instead of smaller, custom models? Like Sonnet 5.5 in this case.
> At least the human has the mouse’s wellbeing at heart.
You don’t know this to be true
If only we had accurate models they could train on.