11 heuristics for choosing (alignment) research projects
I recently spoke with Dane Sherburn about some of the most valuable things he learned as a SERI-MATS scholar.
Here are 11 heuristics he uses to prioritize between research projects:
Impact: Can I actually tell myself a believable story in which this project reduces AI x-risk? (Or better yet; can I make a guesstimate model that helps me estimate the microdooms averted from this project?)
Clarity of research question: Can I easily explain my core research question in a few sentences?
Relevance of research approach: Will my research project actually help me reduce uncertainty on my research question? When I imagine the possible results, are there scenarios where I actually update? Or do I already know (with high probability) what I’m likely to learn?
Mentorship: Would my mentor be able to give me meaningful guidance on this project? If not, would I be able to find one who could?
Feedback loops: Will I be able to get feedback within the first week? First day? Will I have to wait several weeks or months before I know if things are working?
Efficiency: How efficiently will I be able to collect information or run experiments? Will I need to spend a lot of time fine-tuning models? Is there a way to do something similar with pretrained models, so I can run experiments 10-100X more quickly?
Resources: WilI this project need datasets? Large models? Compute? Money? How likely is it that I’ll get the resources I need, and how long will it take?
Excitement: How much does the project subjectively excite me? Do I feel energized about the project?
Timespan: How long would it take to do this project? Would it fit into a window of time that I’m actually willing to devote to it?
Downsides/capabilities externalities: To what extent does the project have capabilities externalities? Could it increase x-risk?
Leaveability: How easy would it be to leave this project if I realize it’s not working out, or I find something better?
Man I really like how short this post is.
Re: 1: Do Dane’s Guestimate models ever yield >1 microdoom estimates for solo research projects? That sounds like a lot.
IIRC Linch estimated in an EA Forum post that we should spend up to ~$100M to reduce x-risk by 1 basis point, i.e. ~$1M per microdoom. Maybe nanodooms would be a better unit.
If your efforts improve the situation by 1 nanodoom, you’ve saved 8 people alive today.
This seems like great advice, thanks!
I’d be interested in an example for what “a believable story in which this project reduces AI x-risk” looks like, if Dane (or someone else) would like to share.