I’m confident that I want to pursue a life of research. But what field do I want to do research in? There are a couple of somewhat closely related, but also fairly different, fields that I’m interested in exploring in 2024-2025 before choosing one and running with it. Of course there is the possibility of changing later, but after a point the EV of choosing one field randomly to dive into becomes higher than the EV of being on the fence.
Q: What are the “live” options for types of research?
Be specific.
pure tcs
I think fundamental tcs research does a pretty good job of points 2&3 of point of research (finding fundamental truths and being intellectually engaging), but in general the pure stuff has less immediate applications.
What specifically would I do here? Some ideas:
- data structures
- distributed algorithms
Plan
- I already have a pretty good picture of what this option looks like, and I basically have the qualifications. So exploring this option can temporarily be put on hold at relatively small cost.
- However, I’d like to do a UROP on distributed algorithms or data structures.
- I would need to decide who I’d want to work with as a grad student and whether to MENG.
Concerns
- I don’t know how to come up with an important research agenda here.
alignment
What specifically would I do here? I’m not sure yet. I think it is somewhat infeasible for me to completely in vacuum come up with my own research agenda here. Instead I think the optimal approach is to learn some things, read some things, and then try to get some mentors who can propose some interesting projects for me to work on. Although, working on open problems from published papers is a reasonable thing to do as well. Maybe something worth noting is that the people at the positions that sound the most exciting to me (e.g., ARC) basically all have PHDs. And as a grad student you probably have some freedom to pursue a wide variety of research agendas. So, the degree to which this option is mutually exclusive from the TCS option is a little bit gray.
Plan
- Read more about current approaches. I’m most excited by ELK ELK1 ELK2 ELK3 ELK4
- Learn stuff. I’ll take DL + RL at MIT. Arena 2 Arena 1 Arena 0 also is a pretty good resource for getting some exposure.
- Do ML UROP.
- Would be good to get a summer job vaguely related to this.
- MSR/google would be cool.
- Worth checking out places like redwood / CHAI / MATS / anthropic
Concerns
- I’m not yet super skilled in this area, so have limited ability to get cool positions.
- It might not be as aesthetically cool as TCS
healthcare / AI for non-alignment-related impact
What specifically would I do here?
- Well, the great thing about not being in the field yet is that I have no attachments.
- So, all other things equal I’ll ask around / research to see what path here seems the most helpful for increasing health / QoL. (i.e., optimizing point 1 of point of research)
Plan
- Do a urop.
Concerns
- Feels like this is hedging “half the impact, half the fun”.
- But really if the claims of AI xrisk are correct then it’s really not so much impact.
- This is like running the dishwasher when the house is on fire: sure, mom will be happy that I did the dishes, but not actually because the house is just going to burn down.
- Maybe this is all just a way of saying, “this feels less live”