Fairly recently I would have strongly endorsed the following claim (e.g., wrt careers):
Conjecture: I can hack myself to feel excited about doing lots of things.
The rational is, feeling good about doing something is plausibly mostly a function of being good at that thing and acting in a way thatās consistent with being excited about it. With a little bit of practice / experience, you can get good at lots of things. It probably also helps if you feel that what youāre doing is important.
Iāve now lost some confidence in this claim. At the very least itās harder than Iād expected to be excited about something new, and hard to want to be.
The math is a bit confusing here. For instance, if I thought that trying to work on important things multiplied your expected positive impact on the world by 1000x compared to not trying, all else equal, but that Iād be 50% less āproductiveā working on something that I didnāt enjoy as much, then thatād still be 500x expected value. But maybe the value from trying to do good is heavy tailed, so the numbers are off. And maybe the numbers are just made up.
I donāt think I have anything particularly good to say about this topic. I guess my 3 cents are:
- The choice between doing good things and doing fun things feels like it should be a false dichotomy.
- Iām extremely grateful that there are some people that have chosen to do good, even if it was hard and required re-thinking their conception of self. For instance, Paul Christiano at one point in time was an MIT student who enjoyed math + TCS and was planning to do TCS research because itās what he was good at. Heās gone on to do a lot of really different stuff, and I suspect heās personally responsible for boosting Pr(future is good) by >1%. I think a similar thing can be said about most people that are trying to make the AI situation go well ā they arenāt āsaintsā that only desire to do good[^1]. Theyāre often people that had some prior intuitions about how theyād spend their life, but have radically updated based on reasoning, even when emotions might discourage them to break from the status quo. Theyāre willing to take high personal risk actions, because they are clear about what their objective function is, and have decided to optimize it.
- My current plan for how I want to spend my life (locally) is as follows:
- For the next 5-6 months Iām going to upskill (including doing object-level useful work when possible, although this seems tricky on my own) for working at ARC.
- I could hedge against the ARC plan working out. Iām not planning to invest much time into this however (beyond some simple things). This decision should be suspected heavily of arising from motivated reasoning. Even so, I feel somewhat comfortable with this decision. In particular, I think that Iām not even close to saturating the āuseful skillsā that itād be good to have for pursuing ARCās agenda. Learning things like inference would clearly be helpful, and having more experience proposing my own research questions and pursuing them would be useful. If you do theoretical computer science and are interested in collaborating on a project, then letās set up a time to brainstorm some ideas. Ideally Iād like a fairly well-defined question with room to branch out with a deliverable of some nice writeup that goes on ArXiv and is maybe published.
- My non-main-line plan is as follows:
- Do TCS research, and get quite good at it.
- Be open to the possibility of doing TCS research that is object-level useful for AI safety, possibly even with some empirical component.
- Be open to trying other weird paths, such as a career in ācommunicationā.
- Communicate more about important issues in the world, and continue to think about problems with the world
- EDIT: actually, Iāve recently been feeling more optimistic that I might be able to do something good even if main-line plan doesnāt pan out. But I think itās okay to figure that out once/if the need arises. And Iām actually fairly optimistic about the main-line strategy.
- Anyways, I donāt claim that this is a very good plan but now that itās articulated it can be refined.