Hi everyone! I am an italian fourth-year medical student who recently discovered the field of AI interpretability through the Effective Altruism community, and I am writing to ask for your advice.
I was originally interested in doing pure neuroscience research, but reading more about this field, in particular the paper 'Zoom In: An Introduction to Circuits' by Chris Olah, made me realise that the kind of study of complex systems I dream of is far more alive in AI interpretability than in the biology-constrained neuroscience I encountered in medical school. This has led me to seriously consider transitioning into this field after graduating, roughly two years from now. While I know resources exist to learn and I am keen on starting, I have deliberately held off. I want a clearer sense of whether someone with my background— no experience in advanced mathematics or programming — has a realistic path before investing serious time. If you think this is the case, what might this path look like? Also, do you see significant career risks in attempting such a transition? If so, would you still recommend trying, or are there alternatives worth considering? Thank you in advance
Hey Valerio,
It's very much possible. What I would say is that you should 1. be realistic about the tradeoffs, 2. Think about your personal fit for AI, and 3. think about what version of this transition is best for you.
Tradeoffs.
Career progression. To start with, as with any career switch, you'll be stepping back a couple of years in seniority in order to make it. For the next few years, you'll obviously find it harder to get a job in AI than a job in medicine, even if you study hard. If you want to be a professor, your chance of doing this in AI is simply much lower than your chance of doing this through medicine +/- neuroscience, because you are currently much stronger at the latter (vis CS students who may be writing AI papers since the middle of undergrad), and furthermore, the latter is less competitive, due to medicine being a largely professional field.
What AI roles look like. In the long-run, working in AI can give you a perfectly competitive salary, compared to medicine, and in some cases a higher one. But the very best roles tend to be in the US, or sometimes the UK, requiring visa applications and waiting periods that can be extremely inconvenient. Also, they are often less secure than a career in medicine. You should also be aware of a common blind spot: as people get older, they care more about the security of their job, and living near their family, or living near where they grew up.
Clear advantages. AI research is far more interesting than being a doctor. The work is much more varied, and people work somewhat less hard. And if you are interested in existential risk, then obviously AI safety research can potentially do something to mitigate this risk in a way that medicine absolutely can't.
Personal fit
The variance in outcomes in an AI career is much more than in medicine. In medicine, it helps a little to be smarter, more conscientious, and more senior, of course. Compared to that, you will notice that in AI (and tech generally), that problem-solving and coding ability is massively important, and can lead to varying outcomes that are way larger than you see in medicine. So before moving into AI, you need to try to be objective and think about it: are you already writing research papers as an undergrad? How is your academic performance? Have you been a mathematics competition kid? Do you want to be in a field that is highly intellectually competitive? Do you find it easy to learn programming, and to build useful demos? Another big element of personal fit is motivation: how much do you care about being involved in interpretability: does it feel like potentially your life cause, or just something that could be nice and interesting. Personally, when I got interested in AI safety, I felt like the former - there was nothing in the world that could convince me away from trying to do something about existential risk. I wanted to try and try until I'd definitively reached a point of failure. If you feel strongly motivated like that, I would definitely recommend you do it, and that's a big consideration.
Thinking through the transition.
If you want to do it, the sooner you transition, the better you are likely to do. The problem with medicine is that there are always a few more years that you can study, to get slightly more money, and slightly more secure, that realistically will take up a lot of your time, and take away a lot of your ability to do outside activities.
I would say the most plausible options are as follows, in decreasing order of urgency:
Although options (2-3) would seem much more secure than (1), they do have their problems. In regard to the save-point, the issue is that I don't know of anyone who actually returned to medicine from EA/AIS work. And the mere option of a high-paying doesn't necessarily make your life all that stable unless you actually do work as a doctor, which will take time. Whereas you could perhaps be making your life more stable by working straight away. So you can argue that (2) it's just wasting time. The issue with (3) is that it's hard to get into a PhD in a strong CS or stats department. The supervisor that you find may be of more of a mixed background, which may be less impressive to AI people. Then again you need to be realistic about where this all ends up: if you become a professor, this is mostly likely to happen at a medical school. If you become a researcher, the most common outcome would not be that you get to work at a frontier lab. Rather, the most lucrative roles that are achievable would more likely be in startups that apply AI to medicine. And, the most impactful and interesting roles that are available are often more research-adjacent, like being a research manager, a policy advisor, or something like that.
Overall thoughts
As you can see, this is not such a simple question to answer. Transitioning from medicine to AI is a complicated judgment. I'm very happy I did it, but for someone in your place, I can only recommend that you understand and weigh the different considerations, and try to arrive at a decision that is best for you.
Thank you very much for your informative answer. I'm glad to know that this transition is doable. It means a lot to me.
I already understood and decided that clinical activity is not for me, so the alternatives would be between research in a neuroscience subject or work in AI interpretability. For research in neuroscience I already have an almost paved road until the PhD, which I could quite easily do at my university. However, at least from the outside, AI interpretability seems to me clearly the best choice in many respects: impact, personal fit (I feel at ease with numbers), kind of work, and raw passion and excitement. I consider AI interpretability my number one interest now, since understanding what's going on in these black boxes is what makes me really curious, and since I'm convinced about the paramount importance of AI safety. If the trade-offs imply a temporary disadvantage because of retraining and initial struggling, it seems definitely a price worth paying. The real brake for me is the possibility of being in some way permanently under the top, permanently among those in the low percentile with respect to contributing to the field, not because I don't have the skills, but because I lost a critical head start. In principle, I would say that if it's a pure gap of knowledge, that gap can at some point be filled, also considering that during high school maths and physics were my strongest subjects by far — but what if there is a snowball effect that permanently favours those with a degree in maths/CS?
Another path that I sort of already discarded, but which may be worth mentioning, is doing a PhD in computational neuroscience as a bridge to AI. Computational neuroscience applies maths and coding to model brain functions. I think it is more accessible for a medical student, since it values medical neuroscience knowledge and is probably less technically demanding, and gives some more basis in maths and coding. The disadvantage is that it is probably less relevant to AI safety than a PhD in AI/ML, and computational neuroscience, while interesting, does not fit into the big picture of how I want to have an impact in the same way as AI does.
It's not a snowball situation. You'll be less and less behind over time, and eventually, you might get to 90% of the aptitude that you would've achieved. Which is fine, except for certain evaluations that happen n years after your PhD - it's going to be extremely difficult to get on any pathway to professorship this way.
As for the issue of how to choose a direction, it's useful to know that you've ruled out clinical activity, and are set on some kind of research. Not everyone would agree with you that interpretability will do much for the paramount risks from AI, but let's take as given that interp is where you want to go. Then a neuroscience PhD that includes wetlab work is going to spend 3-6 years of your life only moving very slowly, and diagonally toward this destination. Comp neuro is not a great idea either, because it's quite remote from your current experience in medicine. There are areas of research that are simultaneously closer to interp, and to medicine. Specifically, researching the interpretability of medical decision-making models, or AI used in medical devices. Basically this kind of stuff: (1) (2) (3). There are probably a dozen CS professors and medical professors who are especially strong at applying modern AI systems and interpretability to medical applications. Consider asking Claude exactly who those professors are. Why not read some of their work? You could try to replicate, or extend one of their studies, or look for an opportunity to do research assisting with one of them, or to study with them.
Thank you again - you've addressed the uncertainties that were concerning me the most.
I was maybe too focused on neuroscience - researching the interpretability of medical decision-making models, or AI used in medical devices, was completely off my radar. I will definitely look into this more and start searching for professors who work in this area.