AI safety scholarships look worth-funding (if other funding is sane)

Fun­ders tend to think that spe­cific sub­sets of early-ca­reer AI safety re­searchers are worth fund­ing:

The pool of AI safety-ori­ented PhD stu­dents across the world is a stronger co­hort in to­tal than any of these par­tic­u­lar groups (be­cause it in­cludes them), and not much weaker on av­er­age. So on the face of it, if those tar­gets are worth fund­ing, then so too should more-gen­eral AI safety re­search schol­ar­ships.

Peo­ple also tend to think broad swathes of early-ca­reer x-risk re­searchers are worth fund­ing:

If AI safety is about as im­por­tant as these other ar­eas, a com­pa­rable amount of tal­ent and su­per­vi­sion is available, then AI safety PhD schol­ar­ships should be similarly worth sup­port­ing.

In­deed, there are as many or more AI safety stu­dents en­ter­ing good pro­grams, and su­per­vi­sors with some in­ter­est in safety like Mar­cus Hut­ter, Roger Grosse, David Du­ve­naud and oth­ers.

On the face of it, stu­dents able to bring fund­ing would be best-equipped to ne­go­ti­ate the best pos­si­ble su­per­vi­sion from the best pos­si­ble school with the great­est pos­si­ble re­search free­dom.

The strongest ap­par­ent ar­gu­ments against are:

  • That PhD schol­ar­ships are ex­pen­sive. In­deed, maybe this isn’t the most effec­tive con­ciev­able fund­ing ob­ject, but it seems as effec­tive as other re­cent pro­jects.

  • Con­cern about ad­verse se­lec­tion of un­funded stu­dents. But this could be miti­gated by mak­ing fund­ing con­di­tional on en­ter­ing a top-10 uni­ver­sity, which would draw a pool of stu­dents that would be stronger than the av­er­age within the field.

  • Con­cerns about the poli­tics of the AI safety and AI field. But this could be miti­gated by pick­ing stu­dents who are su­per­vised to work on top­ics that con­nect to the main­stream.

  • Con­cerns about the amount of time spent by eval­u­a­tors. But the amount of time eval­u­a­tors would spend read­ing ~100 apli­ca­tions (in a year) is pretty small com­pared with the amount of re­search that the ex­tra work ~3 PhD stu­dents se­lected might do over five years.

  • Con­cerns about the vol­ume of ex­cel­lent stu­dents be­ing too low, es­pe­cially af­ter ap­ply­ing some of the filters above. But there are more stu­dents from top schools mov­ing into AI safety than (longter­mist) econ/​philos­o­phy, and GCBRs. Per­haps about as many as the oth­ers put to­gether. Even if half of schol­ar­ships are already funded with grants at the best pos­si­ble school, with the best pos­si­ble su­per­vi­sor, with ad­e­quate free­dom, some will not be, and so there should be room for mul­ti­ple schol­ar­ships per year. Even if the num­ber was less than that, offer­ing some grants would give en­courage­ment to the bright­est up-and-com­ers.

  • Prefer­ence for fund­ing su­per­vi­sors di­rectly, and hav­ing them choose the best PhD stu­dents. But in­ter­ested PhD stu­dents are one of the best ways to get a pro­fes­sor in­ter­ested in a new topic.

This seems like a strong case. Is some­thing be­ing missed?