In Charlie’s cool post about the practical details of getting into alignment work, he notes the very strong correlation between leading alignment research and having a PhD. Sometimes this is naked credentialism (Deepmind, Google Brain), but not always.
Charlie mentions “research taste” as a leading hypothesis for why a PhD is good. Let research taste := “figure out what problem you should try tackling next, explain why that decision makes sense, and repeat this process until you can do it yourself”
This is the main reason I’m in grad school: to get taste through active inference and hundreds of labelled good research decisions. (Also for the optionality and freedom and invites to things.)
But does a PhD improve your taste? Seems clear that it depends entirely on the taste and availability of your advisor. Conditioned on that it seems pretty safe (70%?); unconditioned, it seems way less likely (20%?).
Is this enough for us to bet thousands of years of people’s careers on?
Metascience doesn’t seem to have gotten around to asking if the central mechanism of the field is any good. Here’s one fun methods paper. The obvious thing to do: do it ourselves! Find a lot of people teetering on the edge of accepting offers to grad school and ask them if they want to be randomised! (This design would probably pass an IRB review, because Education Is Good, but morally speaking I’m not sure it should.) This is slow and daft and I challenge you to think of something better, preferably existing evidence.
Since starting a Ph.D. myself, I have updated towards “a Ph.D. is much less useful than I thought” and I usually recommend people not to start one in most instances. However, I think there are some things that a Ph.D. teaches you.
a) Really deeply understand some topic: Spending thousands of hours reading papers, doing some math or coding something means that you are one of a few people globally who have a good understanding of a topic. This can be useful if your topic is useful but also for instrumental reasons. For example, I find it much easier now to dive into a new topic because I feel like it is possible to learn it even if will take some time.
b) Working on your own: This might not be true for every Ph.D. student but for a lot of them. Most of the time, you will work on your own. You will get some supervision and collaborate on some projects but for your first-author papers, you will have to carry the responsibility and do most of the work. During the first year of my Ph.D., I got much more comfortable thinking about a problem even if I couldn’t ask anyone for help. This seems like a good skill when you work on the frontier of a field.
c) A sad but probably true framing: I now think of PhDs as “We throw a smart person at a hard problem and see what happens”. It will almost certainly feel bad and slow and insufficient. But the person will learn a bunch of things that might be valuable. The person might also break and burn out, so it’s a tough trade-off.
d) It’s your only entry to academia: There are a few exceptions but most professors have a Ph.D. If you intend to become a professor, you probably need to do a Ph.D.
The BIG PROBLEM with PhDs (at least in my opinion) is that you can learn most of these skills in other settings as well but with less suffering. Therefore, I would always recommend people to apply to research jobs in industry unless they really want to take the hard route through the Ph.D. If In general, I think you need a strong reason to want to do a Ph.D. and the default should be not doing one even if you intend to work in a research position eventually.
“The BIG PROBLEM with PhDs (at least in my opinion) is that you can learn most of these skills in other settings as well but with less suffering.” Could you elaborate on the suffering part?
https://www.benkuhn.net/grad/