Review of FHI’s Summer Research Fellowship 2020

This post re­views the Fu­ture of Hu­man­ity In­sti­tute’s Sum­mer Re­search Fel­low­ship 2020 in de­tail. If you’re in a hurry, we recom­mend read­ing the sum­mary of ac­tivi­ties, les­sons learned, and com­par­ing costs and benefits sec­tions for a quicker take.

Thanks to Owen Cot­ton-Bar­ratt, Max Daniel, Eli­ana Lorch, Tanya Singh and the sum­mer fel­lows for re­view­ing and im­prov­ing this post.

Sum­mary of activities

  • 27 fel­lows spent 6 weeks on a re­mote fel­low­ship pro­gramme.

    • We got ~300 ap­pli­ca­tions and in­ter­viewed ~50 peo­ple.

    • Back­grounds ranged from un­der­grad­u­ate stu­dents to post­docs.

  • By de­fault, each fel­low had a men­tor and worked on a re­search pro­ject (though there were some ex­cep­tions, and many fel­lows also spent sig­nifi­cant time ex­plor­ing).

    • The men­tors were re­searchers and re­search man­agers at FHI, CSER, OpenPhil, OpenAI, GPI, and MIRI.

    • Ap­prox­i­mately, the pro­jects re­lated to AI gov­er­nance (13 fel­lows), tech­ni­cal AI safety (3 fel­lows), biose­cu­rity (4 fel­lows), macros­trat­egy broadly con­strued (5 fel­lows), and policy (2 fel­lows).

  • This fel­low­ship grew out of the CEA sum­mer fel­low­ship (which CEA wanted to stop run­ning), though was sig­nifi­cantly differ­ent in that a) it was re­mote, and b) we took more fel­lows (in 2019, the CEA fel­low­ship took 11 fel­lows).

Les­sons learned

Meta: we’re stat­ing these con­fi­dently to make clear which di­rec­tion the sum­mer fel­low­ship has up­dated us in. Ob­vi­ously this was just one sum­mer pro­gramme, and in many cases there might be im­por­tant differ­ences which mean that these take­aways wouldn’t gen­er­al­ise. The take­aways are in rough or­der of im­por­tance ac­cord­ing to us.

  • Re­mote fel­low­ships can provide sub­stan­tial value to fel­lows and or­ga­niz­ers, and are worth con­sid­er­ing where in per­son op­tions are not available, or are es­pe­cially costly.

    • Ini­tially we put some prob­a­bil­ity on the whole fel­low­ship be­ing awk­ward and un­pro­duc­tive, but in fact fel­lows had good ex­pe­riences and some good things seem to have come from the pro­gramme. Re­mote­ness also al­lowed us to take more fel­lows, and in the par­tic­u­lar case of Covid-19, also came with lower op­por­tu­nity costs for many fel­lows and or­ganisers.

      • In spite of re­mote­ness, it was pos­si­ble to cre­ate a pos­i­tive, friendly and open cul­ture. Many fel­lows re­marked on this and ap­pre­ci­ated it, and we guess that it helped to in­crease en­gage­ment, sup­port those ex­pe­rienc­ing difficul­ties, and deepen fel­lows’ ex­pe­riences. Model­ling open and au­then­tic com­mu­ni­ca­tion on the part of the or­ganisers seems to have been a ma­jor con­tribut­ing fac­tor to this cul­ture.

    • That said, we still think that there are many ways in which an in per­son fel­low­ship would have been bet­ter, par­tic­u­larly in terms of cul­ture, net­work­ing and some kinds of lo­gis­tics.

  • There is more in­ter­est than we ex­pected in men­tor­ing peo­ple amongst re­searchers in a range of differ­ent EA or­gani­sa­tions. We suc­cess­fully found 18 men­tors for 24 fel­lows (3 fel­lows didn’t have for­mal men­tors for var­i­ous rea­sons), and we con­tacted at least a fur­ther 10 peo­ple who we think would have men­tored if there had been a fel­low work­ing on the right topic. We also didn’t do a com­pre­hen­sive search for men­tors, and in­stead reached out ad hoc to peo­ple in our net­work, so we ex­pect we would have found more in­ter­est if we had searched a bit harder.

  • There were more promis­ing ap­pli­cants for this pro­gramme than we ex­pected. In 2019 the CEA fel­low­ship re­ceived ~90 ap­pli­ca­tions, and this fel­low­ship re­ceived ~300. We think that the num­ber of strong ap­pli­cants was cor­re­spond­ingly larger.

  • Minor difficul­ties with men­tal health were com­mon, and more se­ri­ous ones not ex­tremely rare (al­though this is con­founded by re­mote­ness and Covid-19). 1-1s and other con­ver­sa­tions seemed to help sup­port peo­ple with this. We think it would be good for fu­ture or­ganisers to in­vest more time in think­ing about how to miti­gate the risks from a) im­poster syn­drome, and b) the open ended na­ture of the fel­low­ship and the lack of clear suc­cess met­rics.

  • The per­sonal fit be­tween men­tor and fel­low might po­ten­tially be very im­pact­ful. In the two cases where the fel­low­ship led di­rectly on to fur­ther re­mu­ner­ated col­lab­o­ra­tion be­tween the fel­low and the men­tor, we se­lected the men­tor­ship pairing in part be­cause we ex­pected the two peo­ple to get on well to­gether per­son­ally. It’s un­clear how im­por­tant this was, but we think it’s worth pay­ing at­ten­tion to in fu­ture pairings.

  • A sig­nifi­cant part of the value of (this) fel­low­ship came from pro­vid­ing men­tor­ship and net­work.

  • There’s sig­nifi­cant in­ter­est in the fel­low­ship be­ing longer than 6 weeks, from fel­lows and men­tors, though it’s not clear to us if this is net worth it.

  • Many fel­lows thought that a men­tor with more sub­ject ex­per­tise would have been able to sup­port them more. How­ever, most of these fel­lows also had a strong pos­i­tive ex­pe­rience with their ac­tual men­tor, sug­gest­ing that this is a nice to have rather than an es­sen­tial.

  • Low-in­put at­tempts to cre­ate good on­line office space can be net harm­ful. We set up a de­fault re­mote office space, which ham­pered some more en­er­getic at­tempts from fel­lows to cre­ate good on­line en­vi­ron­ments.

  • Put­ting on talks es­pe­cially for fel­lows was high time cost and only some­what valuable. Espe­cially in the con­text of a re­mote pro­gramme or in the case of recorded talks, we ex­pect that it would have been bet­ter to ei­ther not do this at all, or to in­vest even more time in mak­ing the se­ries re­ally good.

  • Com­mu­ni­cate clearly to fel­lows be­fore the start of the pro­gramme that there are mul­ti­ple differ­ent out­comes that look like suc­cess, if this is the case. We com­mu­ni­cated this at the start of the pro­gramme, which we think was bet­ter than not com­mu­ni­cat­ing it at all or com­mu­ni­cat­ing it later, but we think that com­mu­ni­cat­ing ear­lier than this would likely have bet­ter helped fel­lows to plan and ori­ent.

  • The learn­ing value of run­ning such a pro­gramme for or­ganisers can be very high. Rose had already co-ran a fel­low­ship similar to this one, but still learned a lot (in part be­cause this fel­low­ship was re­mote, in part be­cause she had a more em­pow­ered role on this fel­low­ship). Eli­ana hadn’t pre­vi­ously run a fel­low­ship like this and also learned a lot.

    • Both Rose and Eli­ana had a lot of vi­sion for this fel­low­ship, but not for a re­mote ver­sion of it, and were ini­tially un­ex­cited about a re­mote ver­sion. We were sur­prised that it went well given that we didn’t have a clear vi­sion of how to run this re­motely.

Goals of the fellowship

The main goal of the fel­low­ship is to help young peo­ple to do bet­ter things from a longter­mist per­spec­tive than they would oth­er­wise have done.

We ex­pect most of this change to hap­pen through:

  • Chang­ing fel­lows’ ca­reer plans, by putting fel­lows in the way of con­crete op­por­tu­ni­ties, shift­ing fel­lows’ mo­ti­va­tions, and helping fel­lows to test their fit.

  • Im­prov­ing fel­lows’ net­works, both among other fel­lows and with more se­nior peo­ple.

  • Im­prov­ing fel­lows’ knowl­edge and skills in a range of ways, in­clud­ing re­search taste, re­search ex­pe­rience, and psy­cholog­i­cal em­pow­er­ment.

We try to meet this goal by:

  • Select­ing high po­ten­tial young peo­ple where we think the fel­low­ship is bet­ter than what they could coun­ter­fac­tu­ally be spend­ing time on.

  • Bring­ing them into a rich longter­mist en­vi­ron­ment.

  • Giv­ing them the af­for­dance to work on a rele­vant pro­ject.

  • Pro­vid­ing them with a men­tor who can give them feed­back.

  • In­tro­duc­ing them to each other and rele­vant re­searchers in the space.

De­tail on activities

Timeline of activities

  • Jan­uary/​Fe­bru­ary 2020: Rose and Eli­ana de­cided to take on the pro­ject together

  • 19th Fe­bru­ary: ap­pli­ca­tion page goes live (dead­line 22nd March)

    • We asked for a CV and a re­search proposal

      • The re­search pro­posal was for the pur­poses of eval­u­a­tion; peo­ple were free to work on other things dur­ing the programme

  • March-April: desk re­jects and in­ter­view decisions

  • Late April: interviews

  • Early May: offers

  • 13th July: fel­low­ship started

  • 24th Au­gust: fel­low­ship ended

Sum­mer fellows

There were 27 fel­lows in all. Below is a list of those fel­lows who ex­plic­itly gave us per­mis­sion to name them in this post. See here for more ex­tended bios.

NameBack­groundFo­cus dur­ing the sum­mer
Christo­pher McDon­aldPhysics un­der­grad, now statis­tics mas­ters3D chip­s
Flo­rian DornerMasters in maths, now mas­ters stu­dent in sci­ence, tech­nol­ogy and poli­cyMea­sur­ing Progress in Deep Re­in­force­ment Learn­ing Sam­ple Effi­cien­cy
Ge­or­giana Gil­gal­lonPhilos­o­phy un­der­gradVar­i­ous top­ics in philos­o­phy and macros­trat­e­gy
Gus­tavs Zil­galvisPhysics un­der­gradVar­i­ous top­ics in tech­ni­cal AI safe­ty
Jenny XiaoIR PhD stu­den­tDual moral obli­ga­tions and in­ter­na­tional co­op­er­a­tion
Joel ChristophGlobal poli­tics and eco­nomics mas­ters stu­den­tThe semi­con­duc­tor tal­ent land­scape
Jonas Sand­brinkMed­i­cal doc­tor in train­ingBiose­cu­rity risks as­so­ci­ated with vac­cine tech­nol­ogy; eco­nomic in­cen­tives for novel vac­ci­nes
Joshua Mon­radHealth policy, plan­ning and fi­nanc­ing mas­ters stu­den­tE­co­nomic in­cen­tives for novel vac­ci­nes
Ju­lia VooPoli­ti­cal sci­ence PhD stu­dent, work­ing in tech poli­cyThe re­la­tion­ship be­tween in­ter­na­tional stan­dards and AI
Kwan Yee NgIR un­der­grad and AI gov­er­nance re­searcher/​in­de­pen­dent re­searcher, now mas­ters in China stud­iesAI crash pro­ject­s
Me­gan Kin­ni­men­tPhysics mas­ters stu­den­tFac­tors that in­fluence how peo­ple morally value differ­ent be­ings
Mor­gan MacIn­nesPoli­ti­cal sci­ence PhD stu­den­tThe re­la­tion­ship be­tween AI, pop­u­la­tion and mil­i­tary pow­er
Ni­cholas EmeryPoli­ti­cal sci­ence PhD stu­den­tEs­ti­mat­ing an AI pro­duc­tion func­tion
Nithin Si­vadasElec­tri­cal Eng­ineer­ing PhD stu­dent, now a post­doc at NASAThe statis­ti­cal na­ture of threats to his­tor­i­cal civ­i­liza­tion­s
Nuño Sem­pereIn­de­pen­dent re­searcher and fore­cast­erOp­ti­mal al­lo­ca­tion of philan­thropic re­sources
Sid­dhanth Shar­maJu­nior doc­tor, now mas­ters in pub­lic healthMe­tage­nomic se­quenc­ing and pub­lic health surveillance

Eval­u­a­tion of as­pects of the fellowship

The fel­low­ship as a whole

We asked fel­lows the fol­low­ing ques­tions:

  • How much bet­ter or worse do you feel about the sum­mer fel­low­ship than about what you would have done oth­er­wise?

    • Aver­age 8.58 out of 10, min 5

  • If the same pro­gram (same or­ga­niz­ers, same struc­ture etc.) hap­pened next year, how strongly would you recom­mend a friend (with similar back­ground to you be­fore the fel­low­ship) to ap­ply?

    • Aver­age 8.89 out of 10, min 2 (then 7)

We think these are fairly good re­sponses over­all. The min­i­mum re­sponses to both ques­tions came from the same fel­low.

Length of the programme

Fel­lows’ opinions

  • We asked fel­lows ‘What % of the value you’ve got­ten from the fel­low­ship so far would you at­tribute to each week of the fel­low­ship?’

    • The av­er­age re­sponses were 15%, 17%, 16%, 15%, 14% and 16% for weeks 1,2,3,4,5 and 6

    • This sug­gests that re­turns weren’t diminish­ing over the 6 weeks

  • Of the 27 fel­lows, 5 said in the fi­nal sur­vey that they wished the fel­low­ship had been longer. We guess that an ad­di­tional 3 might also have benefit­ted from a longer fellowship

Men­tors’ opinions

  • 2 men­tors sug­gested mak­ing the fel­low­ship longer in future

Our thoughts

  • It’s worth not­ing that the CEA sum­mer fel­low­ship was 6 weeks in 2018, and fel­lows said it was too short. The 2019 CEA fel­low­ship was there­fore length­ened to 8 weeks, and fel­lows again said it was too short. How­ever, the or­ganisers thought that the 2019 fel­low­ship was a suc­cess, and that an ad­di­tional 2 weeks would not have in­creased the im­pact of the pro­gramme much.

  • There’s some ten­sion be­tween the con­tent of the fel­low­ship and where we think the im­pact comes from. Fel­lows un­der­take a re­search pro­ject dur­ing the fel­low­ship; we think that 6 weeks is a lit­tle too short for this, es­pe­cially given the ex­tra time needed to ori­ent to a new en­vi­ron­ment. How­ever, we don’t think that the main value of the fel­low­ship comes from com­pleted re­search out­puts. In­stead, we ex­pect most of the value to come for ca­reer up­dates, net­work­ing and gain­ing knowl­edge and skills. 6 weeks seems enough to cap­ture sig­nifi­cant value here.

  • We think that or­ganisers next year should se­ri­ously con­sider mak­ing the fel­low­ship longer, but we don’t think it’s clear that they should.

Pro­gram components

  • We asked fel­lows ’How valuable did you find the fol­low­ing [pro­gramme com­po­nents]?

  • The op­tions were Harm­ful, ~Neu­tral, Valuable, Very Valuable, and NA

  • As­sign­ing −1 to Harm­ful, 0 to ~Neu­tral, 1 to Valuable, and 2 to Very valuable, the pro­gramme com­po­nents were rated in the fol­low­ing way:

Men­tor­ingAll fel­lows bar 3 were as­signed a re­search men­tor who met with them roughly once a week.


Th­ese seem like the two most im­por­tant parts of the fel­low­ship struc­ture. In other sec­tions of the sur­vey, mul­ti­ple fel­lows stressed how valuable these things were.

See be­low for more de­tails on men­tor­ing.

Set­ting up the 1-1 in­tro­duc­tions was time in­ten­sive, but seems worth it.

In­tros to FHIers Rose fa­cil­i­tated around 80 1-1 in­tro­duc­tions be­tween fel­lows and (pri­mar­ily) FHI re­searchers


1-on-1s w/​ fel­lowsEvery week, fel­lows were paired with one other fel­low to have a 1-1


1-on-1s w/​ Rose/​Eli­anaRose had at least one 1-1 with ev­ery fel­low, and checked in more fre­quently with some. Eli­ana had a range of 1-1s. The con­tent of these 1-1s was usu­ally check­ing in, dis­cussing prob­lems and progress and get­ting to know each other.


This seems valuable but was high time cost. Over­all we would do it again, as it helped us to run the pro­gramme bet­ter and also miti­gated some risks.
So­cialsWe or­ganised a hand­ful of so­cials at the be­gin­ning of the fel­low­ship, and fel­lows then or­ganised ad­di­tional so­cials them­selves.


This was scored rel­a­tively highly even though we didn’t put much or­ganis­ing effort into mak­ing so­cials hap­pen. 8 fel­lows com­mented that they would have liked more so­cials, so it’s prob­a­ble that there was more value to be had here, though we’re un­sure if this would have been worth ad­di­tional time in­vest­ment on the part of the or­ganisers.
Reg­u­lar FHI talks and sem­i­narsFel­lows were in­vited to most FHI sem­i­nars and re­search events (1-2 per week).


Th­ese talks were hap­pen­ing any­way so it was very cheap to in­vite fel­lows to them.
Open­ing ses­sionOn the first day of the fel­low­ship we held two 1.5 hour open­ing ses­sions to launch the fel­low­ship (2 be­cause time zones meant that it was not pos­si­ble for all fel­lows to at­tend at the same time).


Ca­reer plan­ning work­shopsA col­lab­o­ra­tor ran an op­tional ca­reer plan­ning work­shop in the 5th week of the fel­low­ship, which ~8 fel­lows par­ti­ci­pated in.


Worth run­ning, though in­ter­est­ingly two peo­ple rated it as neu­tral.
Mid­way ses­sionAt the end of the third week, we ran two 1.5 hour mid­way ses­sions, to check in with peo­ple and provide an op­por­tu­nity for fel­lows to re­flect and re­set.


The pro­ject re­view tem­plateIn the 5th week we shared an op­tional tem­plate for fel­lows to use to re­view their pro­jects.


This seems worth it as it was such a small time cost. It’s in­ter­est­ing that it was so much more valuable than the pro­ject plan tem­plate.
SRF talk se­riesWe or­ganised around 14 talks dur­ing the fel­low­ship, from re­searchers in the x-risk space. Some were pre-recorded, oth­ers were live.


This took a fair bit of or­ganis­ing, and pos­si­bly this isn’t a high enough rat­ing to jus­tify the cost. We think that putting in more effort to make the talk se­ries re­ally great, or not do­ing it at all, would have been bet­ter in our case. We also think that a large part of why this wasn’t more valuable was that we didn’t provide a good enough for­mat for syn­chronous view­ing of talks, and that had the fel­low­ship been in per­son, the talks may well have been much more valuable with­out any other changes.
The SRF meta docBe­fore the fel­low­ship, we shared a meta doc with fel­lows de­scribing the pur­pose and struc­ture of the fel­low­ship.


Although it’s a rel­a­tively low rat­ing, it seems clearly im­por­tant for fel­lows to have some doc giv­ing con­text and ex­plain­ing the struc­ture.
Fi­nal pre­sen­ta­tion­sIn the fi­nal week of the fel­low­ship, we or­ganised opt-out fi­nal pre­sen­ta­tions for fel­lows. Th­ese lasted 10 min­utes fol­lowed by 5 min­utes for ques­tions, and were open to all fel­lows, men­tors and FHIers. 23 fel­lows gave fi­nal pre­sen­ta­tions.


Given the time cost to fel­lows and or­ganisers, this may not be worth do­ing in fu­ture (or bet­ter for­mats might make it more valuable).
The pro­ject plan­ning tem­plate In the first week we shared an op­tional tem­plate for fel­lows to use to plan their pro­jects.


This seems worth it as it was such a small time cost.
Check in group meet­ingsAll fel­lows were as­signed to a check in group with 2-4 other fel­lows, to meet once a week to dis­cuss progress and any top­ics of in­ter­est. In the sec­ond week, we sur­veyed fel­lows on how their check in groups were go­ing, hop­ing to be able to in­ter­vene where things weren’t work­ing, but few fel­lows re­sponded and we didn’t take any sig­nifi­cant ac­tions.


This score was pul­led down some­what by the re­sponses from one group which didn’t go well (the score would be 0.96 ex­clud­ing that group).

If this is re­peated, it would be good to have a bet­ter way of in­ter­ven­ing on check in groups when they are not work­ing (e.g. by or­ganisers at­tend­ing all groups once, or check­ing in with ev­ery fel­low in­di­vi­d­u­ally).

In spite of the low score, we ex­pect that given the re­mote na­ture of the fel­low­ship and the large num­ber of fel­lows, hav­ing smaller groups to talk with reg­u­larly may have been im­por­tant for cre­at­ing a friendly, sup­port­ive cul­ture. At least two check in groups con­tinued meet­ing af­ter the fel­low­ship, ad­di­tion­ally sug­gest­ing that this was valuable for some fel­lows.

Light­ning talk­sIn the sec­ond week, we ran three light­ning talk ses­sions. Each fel­low had 5 min­utes for in­tro­duc­ing their work and an­swer­ing ques­tions. The talks were open to all fel­lows, men­tors and FHIers.


Prob­a­bly not worth do­ing, but it’s a bit hard to tell. Fel­lows gave longer talks than would have been ideal, and tried to cover a lot of in­for­ma­tion. If we’d been able to bet­ter prompt fel­lows to give re­ally snappy in­tro­duc­tions to the mo­ti­va­tion for their pro­jects, this might have been worth do­ing.
Weekly slack check-in­Every Mon­day all fel­lows posted what they had done last week, what they were plan­ning to do next week, and any prob­lems or wor­ries in Slack.


This was prob­a­bly worth it as it was cheap, one of the only ways we cre­ated sur­face area be­tween fel­lows in differ­ent check in groups, and wasn’t rated nega­tively by any­one.
The dis­cus­sion group spread­sheetA fel­low set up a spread­sheet to co­or­di­nate dis­cus­sions be­fore and dur­ing the fel­low­ship.


This didn’t re­ally take off, but we think that it was worth it in ex­pec­ta­tion, and prob­a­bly worth it over­all (as 7 fel­lows rated it as valuable).
The de­fault on­line officeIn the first week, we ex­per­i­mented with hav­ing a de­fault on­line office for one day. After the ex­per­i­ment we de­cided to im­ple­ment this as a de­fault office through­out the rest of the fel­low­ship.


This seemed costly, as it ham­pered some more en­er­getic at­tempts by fel­lows to co­or­di­nate on­line office spaces. We think that cre­at­ing a re­ally good on­line office space could be very valuable, but that cheap ver­sions may be worse than noth­ing.


For fellows

We asked fel­lows the fol­low­ing ques­tions:

  • Would you like to see your men­tor men­tor­ing other peo­ple in the fu­ture?

    • Aver­age 4.77 out of 5

  • In what ways do you think an­other men­tor might have been able to sup­port you more?

    • 8 fel­lows: more sub­ject expertise

      • NB 7 of these fel­lows mostly scored the ques­tion about their men­tor men­tor­ing other peo­ple 55 (the fi­nal per­son rated it as 35), and 6 rated men­tor­ing very valuable (the other 2 rated it as valuable)

    • 4 fel­lows: more ac­countabil­ity/​man­age­ment/​structure

    • 2 fel­lows: more ac­tive men­tor­ship in general

    • 1 fel­low each:

      • Feed­back on writ­ten work

      • Differ­ent think­ing style to me

      • More di­rec­tive/​force­ful on re­search directions

      • More introductions

We think that sub­ject ex­per­tise is a nice to have prop­erty in a men­tor, but not strictly nec­es­sary for valuable men­tor­ing pairs to be formed.

For mentors

We asked men­tors the fol­low­ing ques­tions:

  • How much bet­ter or worse do you feel about your men­tor­ing than about what you would have done oth­er­wise? (Count­ing both benefits to your mentee(s) and to you.)

    • Aver­age 7.2 out of 10 (min 6)

  • If the same pro­gram (same or­ga­niz­ers, same struc­ture etc.) hap­pened next year, how strongly would you recom­mend a friend (with similar back­ground to you be­fore the fel­low­ship) to be a men­tor?

    • Aver­age 7.5 out of 10 (min 5)

  • I’ve learned a lot about how to men­tor re­searchers.

    • Aver­age 3.5 out of 5 (min 2)

  • I’m glad that I vol­un­teered to be a men­tor.

    • Aver­age 4.3 out of 5 (min 3)

  • If we ran the fel­low­ship again, what would you recom­mend we change?

    • 2 men­tors men­tioned pro­vid­ing more ad­vice and prompt­ing on mentoring

    • 1 men­tor each mentioned

      • Shar­ing tem­plates for men­tors to re­quest feed­back from mentees

      • More in­volve­ment of men­tors in the se­lec­tion process

      • Nudg­ing men­tors to make introductions

      • Up­dat­ing men­tors on the ca­reer plan­ning ses­sion we ran

In one sense, the rat­ings men­tors gave seem quite low. How­ever, they clear the bar of ‘worth my time over­all’.


We asked fel­lows ‘Do you have any com­ments on the cul­ture of the Sum­mer Re­search Fel­low­ship?’

  • 12 fel­lows com­mented on the cul­ture be­ing quite good to fan­tas­tic in a generic way

  • 9 fel­lows men­tioned the cul­ture be­ing warm/​kind/​friendly

  • 3 fel­lows ap­pre­ci­ated the open­ness of the cul­ture (in terms of shar­ing strug­gles and difficul­ties)

  • 1 fel­low men­tioned each of the fol­low­ing as a prob­lem with the cul­ture:

    • Lack of Slack engagement

    • Shar­ing cu­ri­osi­ties not be­ing suffi­ciently common

    • In­suffi­cient steps taken to miti­gate im­poster syndrome

    • Too much of an ‘in’ cul­ture around EA/​rationality

  • Fel­lows also com­mented on what had helped con­tribute to the culture

    • 7 fel­lows men­tioned mod­el­ling from Rose and Eliana

    • 2 fel­lows men­tioned the open­ing session

We think the cul­ture was sur­pris­ingly good for a re­mote fel­low­ship, and we guess that it helped to in­crease en­gage­ment, sup­port those ex­pe­rienc­ing difficul­ties, and deepen fel­lows’ ex­pe­riences.

Men­tal health

  • We asked fel­lows, ‘Did you ex­pe­rience any difficul­ties with men­tal health dur­ing the fel­low­ship?’

    • 15 peo­ple re­ported some difficul­ties, of whom 3 re­ported sig­nifi­cant difficul­ties (ac­cord­ing to us)

  • Of those 15 peo­ple, 4 re­ported that SRF con­tributed to the difficul­ties they ex­pe­rienced, speci­fi­cally through:

    • Open ended com­mit­ments and un­clear paths to suc­cess (2 fel­lows)

    • Feel­ings of im­poster syn­drome (1 fel­low)

    • Lo­gis­ti­cal stuff to do with time zones (1 fel­low, and we know of at least one ad­di­tional fel­low for whom this posed non-triv­ial challenges)

  • 13 fel­lows re­ported that the fel­low­ship sup­ported them in their difficul­ties, mostly via 1-1s and other conversations

    • The in­ter­sec­tion be­tween this group of 13 and the group of 15 who re­ported difficul­ties is 12

    • The in­ter­sec­tion be­tween this group of 13 and the group of 4 where SRF con­tributed to their difficul­ties is 3

  • 2 fel­lows some­times felt unwelcome

    • In one case the fel­low felt un­wel­come in some heav­ily EA con­ver­sa­tions where they had less back­ground than the other participants

    • In an­other case the fel­low felt un­wel­come in some fel­low-or­ganised dis­cus­sion groups

We think that given that the fel­low­ship was re­mote, this rate of men­tal health difficul­ties isn’t sur­pris­ing, and also that the coun­ter­fac­tual is un­clear given the global pan­demic.

We also think that any­one hav­ing difficul­ties with men­tal health that are ex­ac­er­bated by the fel­low­ship is bad, and we should en­deav­our to do bet­ter. In par­tic­u­lar, we think fu­ture or­ganisers should think more about ways of miti­gat­ing the men­tal health risks of the fel­low­ship be­ing open ended and (in its cur­rent form) hav­ing many and there­fore un­clear paths to suc­cess. It’s worth not­ing that we think that the op­po­site ex­treme (of a very clear and par­tic­u­lar goal) is likely to be challeng­ing for other fel­lows, and that the right solu­tion will take vari­a­tion among fel­lows into ac­count.


We cre­ated a sep­a­rate Slack workspace for the fel­lows and their men­tors. Men­tor en­gage­ment with the Slack workspace was op­tional, and in the end, min­i­mal. We did not add fel­lows to the reg­u­lar FHI Slack workspace, as it con­tains po­ten­tially sen­si­tive in­for­ma­tion, and we didn’t want to flood the Slack with new mem­bers.

The Slack worked well as a co­or­di­na­tion tool, but didn’t be­come a rich in­tel­lec­tual space. There weren’t many ob­ject-level re­search dis­cus­sions, or much so­cial en­gage­ment.

Not adding fel­lows to the FHI Slack worked well. We think we still man­aged to in­tro­duce fel­lows to FHIers in a way that was valuable.

Costs of run­ning the fellowship


  • The fi­nan­cial costs were all spent on stipends for fel­lows.

  • Re­searcher time talk­ing to fel­lows. As or­ganisers, we set up roughly 80 1-1 pairings be­tween FHI re­searchers and fel­lows. Prob­a­bly some of these never ac­tu­ally took place, but in ad­di­tion many men­tors made in­tro­duc­tions too. As­sum­ing that 1-1s lasted 60 min­utes on av­er­age, this cost maybe 100 hours of FHI re­searcher time.

  • Ad­min time ad­minis­ter­ing se­lec­tion, set­ting up stipends and schedul­ing events (per­haps 110-130 hours).

  • A small num­ber of peo­ple (2-6) gave us feed­back on our plans, which cost a small amount of time (maybe 4-20 hours over­all).

  • Around 20 FHIers spent some time re­view­ing some ap­pli­ca­tions. We guess this cost 40-60 hours to­tal.

To mentors

  • There were 18 men­tors, who were re­searchers and re­search man­agers at FHI, CSER, OpenPhil, OpenAI, GPI and MIRI.

  • In ad­di­tion, there were 2 in­for­mal men­tors who agreed to speak more oc­ca­sion­ally with a fel­low. Th­ese in­for­mal men­tors were at CSET and GPI.

  • All men­tors were vol­un­teers.

  • Most men­tors met with fel­lows once a week for 30-60 min­utes. Th­ese 18 men­tors had 24 fel­lows be­tween them.

    • Guess­ing that two thirds of the time men­tors spent was liter­ally in meet­ing fel­lows, this makes for a net time cost to men­tors of roughly ~170 hours.

  • NB all men­tors thought men­tor­ing was net bet­ter than their coun­ter­fac­tual (av­er­age 7.2/​10 com­pared with the coun­ter­fac­tual; min­i­mum 610).

To the SRF organisers

  • Eli­ana Lorch and Rose Had­shar ran the fel­low­ship. Nei­ther tracked their time, but we guess that com­bined we spent the fol­low­ing amount of time on the fel­low­ship:

    • Selec­tion: maybe 3 weeks of work time

    • Plan­ning the pro­gramme: maybe 1.5 weeks of work time

    • Run­ning the pro­gramme: maybe 2.5 weeks of work time

  • We feel that this time was bet­ter spent on the fel­low­ship than on our coun­ter­fac­tu­als (in Rose’s case, more in­vest­ment into RSP cul­ture and pro­gramme struc­ture; in Eli­ana’s case, more work on mod­el­ing).

  • This makes for a to­tal of ~280 or­ganis­ing hours. The CEA fel­low­ship in 2019 cost more like ~200 hours to its or­ganisers, so this ver­sion was more time in­ten­sive. We ex­pect that this comes in large part from the fact that we had triple the num­ber of ap­pli­ca­tions, took nearly three times the num­ber of fel­lows, and needed to re­think the pro­gramme sub­stan­tially to make it re­mote.

  • Or­ganis­ing was stress­ful for both Rose and Eli­ana at times.

To fellows

  • 2 fel­lows re­ported feel­ing iso­lated at points dur­ing the fellowship

  • 3 fel­lows re­ported that the fel­low­ship ex­ac­er­bated sig­nifi­cant men­tal health difficul­ties they were facing

    • In one case, the fel­low­ship con­tributed via pres­sure and feel­ings of im­poster syndrome

    • In the other two cases, the fel­low­ship con­tributed via un­clear ex­pec­ta­tions and suc­cess criteria

  • Op­por­tu­nity costs

    • 1 fel­low rated the fel­low­ship as 510 com­pared with their coun­ter­fac­tual and we think had a frus­trat­ing ex­pe­rience of the fellowship

    • All other fel­lows rated the fel­low­ship as net bet­ter than their coun­ter­fac­tual (av­er­age 8.6/​10 com­pared with the coun­ter­fac­tual)

Benefits pro­duced by the fellowship


  • Un­less stated oth­er­wise, the fol­low­ing is based on a mix of (i) re­sponses to feed­back forms all fel­lows and men­tors filled out at the end of the pro­gramme, (ii) Rose’s and Eli­ana’s im­pres­sions based on in­ter­ac­tions with the fel­lows, their men­tors etc.

  • A lot of this is sub­jec­tive.

  • It’s also some­what sen­si­tive as it re­lates to par­tic­u­lar peo­ple. We’ve anonymised as far as pos­si­ble.

  • We think that most of the benefits of the fel­low­ship won’t be visi­ble un­til more time has passed. We in­tend to sur­vey fel­lows again 6 months af­ter the fel­low­ship, and pos­si­bly 2 years af­ter­wards too.

To fellows

Ca­reer updates

  • Con­crete opportunities

    • 1 fel­low found a paid re­search po­si­tion work­ing for the per­son who men­tored them on the fellowship

    • 1 fel­low re­ceived two grants to con­tinue their fel­low­ship pro­ject and to work on re­lated research

    • 2 fel­lows have on­go­ing col­lab­o­ra­tion/​men­tor­ship relationships

    • 1 fel­low be­came an af­fili­ate at a longter­mist re­search organisation

    • 3 fel­lows have po­ten­tial fu­ture collaborators

  • Motivations

    • 4 fel­lows be­came more likely to work on things with longter­mist implications

  • Studying

    • 4 fel­lows up­dated to­wards fur­ther study (with longter­mist rele­vance)

  • Re­search updates

    • 4 fel­lows got a clearer idea of what to work on dur­ing their PhD or mas­ters degree

    • 2 fel­lows up­dated against re­search careers

    • 2 fel­lows up­dated to­wards global pri­ori­ties research

    • 1 fel­low shifted their idea of what to ap­ply to do a PhD on

  • Meta

    • 11 fel­lows have men­tors who in­tend to speak with them in fu­ture about their ca­reer plans

      • In one case the men­tor in­tends to sub­stan­tially en­gage in dis­cus­sions around choos­ing a PhD topic


  • The me­dian fel­low re­ported know­ing 5 peo­ple be­fore the fel­low­ship who met the fol­low­ing crite­ria:

    • a) have spent more time than you think­ing about or work­ing on ques­tions of rele­vance to the long-term fu­ture, and

    • b) you’d feel com­fortable reach­ing out to for ad­vice or feedback

  • The me­dian fel­low re­ported know­ing an ad­di­tional 6 peo­ple as a re­sult of the fel­low­ship who met the same criteria

  • This rep­re­sents an av­er­age in­crease of fel­lows’ long ter­mist net­works of around 100%

  • One fel­low who had very lit­tle ex­po­sure to EA prior to the fel­low­ship re­ported that it was a pos­i­tive in­tro­duc­tion to EA for them

Knowl­edge and skills

  • 9 fel­lows gained sig­nifi­cant knowl­edge on var­i­ous as­pects of longtermism

  • 8 fel­lows gained sig­nifi­cant re­search skills

    • In most cases, these were fel­lows with lit­tle prior re­search experience

    • In a few cases, skills were gained by more ex­pe­rienced re­searchers around re­search taste

  • 4 fel­lows re­ported im­prov­ing on other skills or traits

    • Agenti-ness, pro­duc­tivity, and con­fi­dence were named


This list con­tains out­puts where there was a pub­li­cly ac­cessible link, and/​or where fel­lows were happy to in­clude their name and email ad­dress against their work. There were many other out­puts where nei­ther of these things ap­plied, which we have ex­cluded from this list.

Note that in some cases only parts of the out­put were pro­duced dur­ing the fel­low­ship. How­ever, we think that for most if not all of these, these texts would not ex­ist, or would have been pub­lished later or with worse qual­ity, if not for the fel­low­ship.

Com­pleted out­puts:

Notable works in progress in­clude:

Other effects

To mentors

  • 1 men­tor’s fel­low was work­ing on a pro­ject from the men­tor which oth­er­wise the men­tor would have wanted to do them­selves. The fel­low’s time was a bit un­der 1:1 equiv­a­lent of the men­tor’s time on the pro­ject, and the fel­low ex­tended the pa­per fol­low­ing on from the fellowship

  • 3 men­tors up­dated to­wards men­tor­ing in fu­ture, sig­nifi­cantly in 2 cases

  • 2 men­tors clar­ified some of their own re­search thinking

  • 4 men­tors learned some use­ful ob­ject level things

  • 1 men­tor re­ported im­prov­ing at mentoring

  • Men­tors up­dated on how promis­ing 9 fel­lows were in the course of in­ter­act­ing with them

Com­par­ing costs and benefits

We think that the fel­low­ship was pretty clearly net pos­i­tive, but that it also wasn’t clearly an out­stand­ing suc­cess.

In par­tic­u­lar, at the end of the 2019 CEA fel­low­ship, we think there were pro­por­tion­ally more clear suc­cess cases to point to than there are at the same point for the 2020 FHI fel­low­ship. This might in part be be­cause there were fewer fel­lows in 2019, which might mean a) fel­lows had more in­put per head and this cre­ated more value, and b) or­ganisers had a bet­ter sense of how in­di­vi­d­ual fel­lows had de­vel­oped, mak­ing it eas­ier to spot suc­cesses. It might also be be­cause the fel­low­ship was re­mote and this made it less valuable. Small sam­ple size and fre­quency of out­liers might also mean that this is ex­plained just by luck.

We ex­pect more of the im­pact of this fel­low­ship to be­come clear later as fel­lows progress in their ca­reers, and in­tend to sur­vey fel­lows again 6 months af­ter the fel­low­ship and pos­si­bly 2 years af­ter.


We’re sure we made other mis­takes too, but the ones that stand out in chronolog­i­cal or­der are:

  • Rose in­vested in­suffi­cient time in man­ag­ing the per­son pro­vid­ing ad­minis­tra­tive sup­port dur­ing the ap­pli­ca­tion pro­cess. This cre­ated un­nec­es­sary in­effi­ciency, con­fu­sion and de­lays for or­ganisers and ap­pli­cants.

  • We’d make some hiring de­ci­sions differ­ently now, and we think for some of them we could have re­al­ized this ex ante.

  • We didn’t com­mu­ni­cate promptly enough with fel­lows about the ex­pec­ta­tions we had of them, or the struc­ture and sched­ule of the pro­gramme.

  • In the first week the timings of a num­ber of events were changed last minute, and this cre­ated con­fu­sion that seemed par­tic­u­larly costly at the be­gin­ning of the fel­low­ship.

  • We in­sti­tuted a de­fault on­line office space, but didn’t in­vest any en­ergy or thought into in­te­grat­ing this well with other parts of the fel­low­ship or fel­lows’ ini­ti­a­tives. This led to a fairly in­ac­tive on­line office space, and prob­a­bly hin­dered some more en­er­getic ini­ti­a­tives from fel­lows.

  • We didn’t com­mu­ni­cate clearly with fel­lows about how to pre­pare and give good light­ning talks, and this made these ses­sions less valuable than oth­er­wise.

  • We didn’t put con­scious at­ten­tion on how to elicit nega­tive feed­back (anony­mous feed­back op­tions, op­tions to give feed­back to non-or­ganisers). We ex­pect this leads to a pos­i­tive skew in our eval­u­a­tion here.

We may con­sider shar­ing some ma­te­ri­als and re­sources with or­ganisers of similar pro­grammes in fu­ture. If you think these ma­te­ri­als would be use­ful to you, please email srf@philos­o­ ex­plain­ing why this would be use­ful, and we’ll con­sider your re­quest.