Jonas Vollmer

• Jonas Vollmer 18 Apr 2019 9:25 UTC

  4 points

  on: Sup­port­ing global co­or­di­na­tion in AI de­vel­op­ment: Why and how to con­tribute to in­ter­na­tional AI standards

  This is great! I think it could be worth em­pha­siz­ing more that you’re es­sen­tially mak­ing a linkpost for the FHI /​ GovAI tech­ni­cal re­port Stan­dards for AI Gover­nance: In­ter­na­tional Stan­dards to En­able Global Co­or­di­na­tion in AI Re­search & Devel­op­ment.

• Jonas Vollmer 11 Apr 2019 7:39 UTC

  4 points

  in reply to: Michelle_Hutchinson's comment on: Long Term Fu­ture Fund: April 2019 grant decisions

  I tend to think that the net­work con­straints are bet­ter ad­dressed by solu­tions other than ad-hoc fixes (such as more proac­tive in­ves­ti­ga­tions of grantees), though I agree it’s a con­cern and it up­dates me a bit to­wards this not be­ing a good idea.

  I wasn’t sug­gest­ing de­cid­ing the op­por­tu­nity cost case by case. In­stead, grant eval­u­a­tors could as­sume a fixed cost of e.g. $2k. In terms of es­ti­mat­ing the benefit of mak­ing the grant, I think they do that already to some ex­tent by pro­vid­ing nu­mer­i­cal rat­ings to grants (as Oliver ex­plains here). Also, be­ing aware of the $10k rule already cre­ates a small amount of work. Over­all, I think the ad­di­tional amount of work seems neg­ligibly small.

  ETA: Set­ting a lower thresh­old would al­low us to a) avoid turn­ing down promis­ing grants, and b) re­move an in­cen­tive to ask for too much money. That seems pretty use­ful to me.

• Jonas Vollmer 10 Apr 2019 14:00 UTC

  16 points

  in reply to: Ben_Kuhn's comment on: Long Term Fu­ture Fund: April 2019 grant decisions

  I ac­tu­ally think the $10k grant thresh­old doesn’t make a lot of sense even if we as­sume the de­tails of this “op­por­tu­nity cost” per­spec­tive are cor­rect. Grants should fulfill the fol­low­ing crite­rion:

  “Benefit of mak­ing the grant” ≥ “Fi­nan­cial cost of grant” + “CEA’s op­por­tu­nity cost from dis­tribut­ing a grant”

  If we as­sume that there are large im­pact differ­ences be­tween differ­ent op­por­tu­ni­ties, as EAs gen­er­ally do, a $5k grant could eas­ily have a benefit worth $50k to the EA com­mu­nity, and there­fore eas­ily be worth the $2k of op­por­tu­nity cost to CEA. (A po­ten­tial jus­tifi­ca­tion of the $10k thresh­old could ar­gue in terms of some sort of “mar­ket effi­ciency” of grant­mak­ing op­por­tu­ni­ties, but I think this would only jus­tify a rigid thresh­old of ~$2k.)

  IMO, a more de­sir­able solu­tion would be to have the EA Fund com­mit­tees fac­tor in the op­por­tu­nity cost of mak­ing a grant on a case-by-case ba­sis, rather than hav­ing a rigid “$10k” rule. Since EA Fund com­mit­tees gen­er­ally con­sist of smart peo­ple, I think they’d be able to un­der­stand and im­ple­ment this well.

• Jonas Vollmer 10 Apr 2019 13:59 UTC

  5 points

  in reply to: Habryka's comment on: Long Term Fu­ture Fund: April 2019 grant decisions

  (moved this com­ment here)

• Jonas Vollmer 10 Apr 2019 13:47 UTC

  2 points

  in reply to: Greg_Colbourn's comment on: $100 Prize to Best Ar­gu­ment Against Donat­ing to the EA Hotel

  I agree. The changes you’re mak­ing seem great! I also like the con­cise de­scrip­tion.

  (Will get back on some of the de­tails via email, e.g., not sure 95% CIs are worth the effort.)

  (Strong up­voted.)

• Jonas Vollmer 8 Apr 2019 14:39 UTC

  10 points

  in reply to: Habryka's comment on: Long Term Fu­ture Fund: April 2019 grant decisions

  On #3, this goes in a similar di­rec­tion.

• Jonas Vollmer 8 Apr 2019 7:03 UTC

  9 points

  in reply to: Halffull's comment on: How x-risk pro­jects are differ­ent from startups

  Do you think in­ten­tional in­sights did a lot of dam­age? I’d say it was rec­og­nized by the com­mu­nity and pretty well han­dled whole do­ing al­most no dam­age.

  As I also say in my above-linked talk, if we think that EA is con­strained by vet­ting and by se­nior staff time, things like InIn have a very sig­nifi­cant op­por­tu­nity cost be­cause they tend to take up a lot of time from se­nior EAs. To get a sense of this, just have a look at how long and thor­ough Jeff Kauf­man’s post is, and how many peo­ple gave in­put/​feed­back—I’d guess that’s sev­eral weeks of work by se­nior staff that could oth­er­wise go to­wards re­solv­ing im­por­tant bot­tle­necks in EA. On top of that, I’d guess there was a lot of in­ter­nal dis­cus­sion in sev­eral EA orgs about how to han­dle this case. So I’d say this is a good ex­am­ple of how a sin­gle per­son can have a lot of nega­tive im­pact that af­fects a lot of peo­ple.

  I wasn’t ask­ing for ex­am­ples from EA, just the type of pro­jects we’d ex­pect from EAs.

  The above-linked 80k ar­ti­cle and EAG talk men­tion a lot of po­ten­tial ex­am­ples. I’m not sure what else you were hop­ing for? I also gave a con­cise (but not com­plete) overview in this face­book com­ment.

• Jonas Vollmer 5 Apr 2019 18:32 UTC

  20 points

  on: How x-risk pro­jects are differ­ent from startups

  Also rele­vant: Speed­ing up so­cial sci­ence 10-fold, how to do re­search that’s ac­tu­ally use­ful, & why plenty of star­tups cause harm (and Spencer’s blog post)

  Spencer and Rob think plenty of star­tups are ac­tu­ally harm­ful for so­ciety. Spencer ex­plains how com­pa­nies reg­u­larly cause harm by in­jur­ing their cus­tomers or third par­ties, or by draw­ing peo­ple and in­vest­ment away from more use­ful pro­jects.

  So EAs should be more cau­tious even about or­di­nary start-ups in com­par­i­son to the typ­i­cal Ven­ture Cap­i­tal­ist.

• Jonas Vollmer 5 Apr 2019 18:24 UTC

  8 points

  in reply to: Halffull's comment on: How x-risk pro­jects are differ­ent from startups

  A few ex­am­ples are men­tioned in the re­sources linked above. The most well-known and com­monly ac­cepted one is In­ten­tional In­sights, but I think there are quite a few more.

  I gen­er­ally pre­fer not to make nega­tive pub­lic state­ments about well-in­ten­tioned EA pro­jects. I think this is prob­a­bly the rea­son why the ex­am­ples might not be salient to ev­ery­one.

• Jonas Vollmer 3 Apr 2019 7:10 UTC

  10 points

  on: Open Thread #44

  EA Fo­rum posts are of­ten hard to find through Google search

  E.g., when search­ing on Google for “Long Term Fu­ture Fund,” this re­sult only shows up at the bot­tom of the sec­ond page of search re­sults, even af­ter many user pro­file pages that seem much less rele­vant. Maybe this some­thing that can be fixed in some ways?

  I guess the solu­tion would be to use the fo­rum’s own search fea­ture, but many peo­ple (in­clud­ing my­self un­til just now) don’t rou­tinely use that fea­ture and pre­fer us­ing Google search.

  Another ex­am­ple: Search­ing on Google for “ran­dom fund­ing max daniel effec­tive al­tru­ism fo­rum” doesn’t lead to this re­sult, you have to search for “ran­dom fund­ing max_daniel effec­tive al­tru­ism fo­rum” in­stead. I wasn’t aware that Google re­sponds to the “un­der­line” char­ac­ter, and it seems strange that it’s so sen­si­tive to it.

  Maybe this is all Google’s mis­take, but I find this a bit hard to be­lieve.

• Jonas Vollmer 1 Apr 2019 16:25 UTC

  6 points

  on: $100 Prize to Best Ar­gu­ment Against Donat­ing to the EA Hotel

  Sup­pose that, on some level of gen­eral com­pe­tence, Alice is 95th per­centile among EAs on the Fo­rum and is work­ing on her own EA pro­ject in­de­pen­dently, while Bob is of 30th per­centile com­pe­tence and is work­ing on his pro­ject while so­cially im­mersed in his many in-per­son EA con­tacts.

  I agree be­ing im­mersed is im­por­tant be­cause risks are hard to an­ti­ci­pate for a sin­gle in­di­vi­d­ual. I would ar­gue that the sce­nario seems some­what ar­tifi­cial, as gen­eral com­pe­tence of some­one not in­ter­act­ing with EAs is un­likely to be in the 95th per­centile.

  How­ever, once these do­main-spe­cific pit­falls are pointed out to you, it’s not that cog­ni­tively tax­ing to grok them and ad­just your think­ing/​ac­tions ac­cord­ingly.

  I agree. How­ever, this is not re­ally about skill or in­tel­li­gence: Hu­mans in gen­eral of­ten don’t take crit­i­cal feed­back nearly as se­ri­ously as they should, and of­ten don’t ad­just their think­ing/​ac­tions due to sunk costs, want­ing to save face in their peer group, grandiose per­son­al­ity, etc. This also ap­plies to EAs (maybe some­what but not vastly less so).

  From look­ing at the pub­lished list of EA Ho­tel res­i­dents, I ten­ta­tively think some peo­ple’s work might come with high down­side risk, while oth­ers have high up­side po­ten­tial and seem worth sup­port­ing. I’m not sure how this bal­ances out. Dis­cussing in­di­vi­d­ual pro­jects in pub­lic seems difficult, which is maybe part of the rea­son why peo­ple find the ar­gu­ments against fund­ing the EA Ho­tel un­con­vinc­ing. All else equal, I’d prob­a­bly pre­fer some­thing like Aaron Gertler’s ap­proach of “look­ing at the Ho­tel’s guest list, pick­ing the best-sound­ing pro­ject, and offer­ing money di­rectly to the per­son be­hind it.” I have also shared some thoughts for how to de­sign the ad­mis­sion pro­cess with EA Ho­tel staff.

  (If one ac­cepted the premise that down­side risk is preva­lent and sig­nifi­cant, one could ar­gue that any dona­tion to the EA Ho­tel that doesn’t set in­cen­tives to re­duce down­side risk might coun­ter­fac­tu­ally re­place a dona­tion that does. I’m not sure this ar­gu­ment works, but it could be worth think­ing about.)

  (All my per­sonal opinion, not speak­ing for any­one here.)

  Edited to add: In many ways, the EA Ho­tel acts like a de facto EA grant­maker, so the con­cerns out­lined in my com­ment here ap­ply:

  When long-ter­mist grant ap­pli­ca­tions don’t get funded, the rea­son usu­ally isn’t lack of fund­ing, but one of the fol­low­ing:

  --The grant­maker was un­able to vet the pro­ject (due to time con­straints or lack of do­main ex­per­tise) or at least thought it was a bet­ter fit for a differ­ent grant­maker.

  --The grant­maker thought the pro­ject came with a high risk of ac­ci­den­tal harm.

  (…)

  High-qual­ity grant ap­pli­ca­tions tend to get funded quickly and are thereby elimi­nated from the pool of pro­pos­als available to the EA com­mu­nity, while ap­pli­cants with higher-risk pro­pos­als tend to ap­ply/​pitch to lots of fun­ders. This means that on av­er­age, pro­pos­als sub­mit­ted to fun­ders will be skewed to­wards high-down­side-risk pro­jects, and fun­ders could them­selves eas­ily do harm if they end up sup­port­ing many of them.

  (…)

• Jonas Vollmer 1 Apr 2019 16:06 UTC

  11 points

  in reply to: John_Maxwell_IV's comment on: $100 Prize to Best Ar­gu­ment Against Donat­ing to the EA Hotel

  Rele­vant: When should EAs al­lo­cate fund­ing ran­domly? An in­con­clu­sive liter­a­ture re­view.

• Jonas Vollmer 22 Mar 2019 20:04 UTC

  2 points

  in reply to: Brendon_Wong's comment on: EA An­gel Group: Ap­pli­ca­tions Open for Per­sonal/​Pro­ject Funding

  Thanks for the thor­ough re­sponse! I think I agree with what you said, and I think the pro­cess you men­tioned seems ad­e­quate to ad­dress the risks (if im­ple­mented well).

  My per­cep­tion is that ap­pli­ca­tion shar­ing could help ad­dress vet­ting con­straints be­cause it al­lows other fun­ders (who may have more ex­per­tise in a par­tic­u­lar area) to help with vet­ting. I think other fun­ders prob­a­bly don’t have rol­ling ap­pli­ca­tions be­cause of the in­creased effort this en­tails, so in that sense rol­ling ap­pli­ca­tions can also help re­solve vet­ting con­straints.

• Jonas Vollmer 22 Mar 2019 15:49 UTC

  2 points

  in reply to: EA person's comment on: EA is vet­ting-constrained

  This post seems very in­sight­ful to me, and it seems like it worked out very well in terms of up­votes (and it seems like it would in­crease your chances of get­ting fund­ing)? I’d be in­ter­ested to learn who wrote this, but of course no need to say if you pre­fer not to. :)

• Jonas Vollmer 22 Mar 2019 14:43 UTC

  30 points

  on: EA An­gel Group: Ap­pli­ca­tions Open for Per­sonal/​Pro­ject Funding

  When long-ter­mist grant ap­pli­ca­tions don’t get funded, the rea­son usu­ally isn’t lack of fund­ing, but one of the fol­low­ing:

  • The grant­maker was un­able to vet the pro­ject (due to time con­straints or lack of do­main ex­per­tise) or at least thought it was a bet­ter fit for a differ­ent grant­maker.

  • The grant­maker thought the pro­ject came with a high risk of ac­ci­den­tal harm.

  This post con­tains great ideas for re­solv­ing the former point, but doesn’t demon­strate high aware­ness of the lat­ter con­cern. Aware­ness of these risks seems im­por­tant to me, es­pe­cially for fun­ders: High-qual­ity grant ap­pli­ca­tions tend to get funded quickly and are thereby elimi­nated from the pool of pro­pos­als available to the EA com­mu­nity, while ap­pli­cants with higher-risk pro­pos­als tend to ap­ply/​pitch to lots of fun­ders. This means that on av­er­age, pro­pos­als sub­mit­ted to fun­ders will be skewed to­wards high-down­side-risk pro­jects, and fun­ders could them­selves eas­ily do harm if they end up sup­port­ing many of them. I’d be in­ter­ested in your thoughts on that.

  I re­ally like that you’re pro­vid­ing feed­back to ap­pli­cants! In gen­eral, I wish the EA com­mu­nity was more proac­tive with pro­vid­ing crit­i­cal feed­back.

• Jonas Vollmer 28 Feb 2019 20:33 UTC

  4 points

  in reply to: JoshYou's comment on: After one year of ap­ply­ing for EA jobs: It is re­ally, re­ally hard to get hired by an EA organisation

  I would ask Open Phil whether they’d be okay with you shar­ing it with the or­ga­ni­za­tion to your ap­ply­ing to (ideally only once you’re past the first stage, and only if the other or­ga­ni­za­tion ex­pressed in­ter­est).

• Jonas Vollmer 27 Feb 2019 23:50 UTC

  4 points

  in reply to: Ben_Kuhn's comment on: After one year of ap­ply­ing for EA jobs: It is re­ally, re­ally hard to get hired by an EA organisation

  Good point, agree it de­pends on the type of work.

• Jonas Vollmer 27 Feb 2019 9:20 UTC

  12 points

  in reply to: Ben_Kuhn's comment on: After one year of ap­ply­ing for EA jobs: It is re­ally, re­ally hard to get hired by an EA organisation

  Two points that speak against this view a bit:

  • It seems eas­ier to in­crease the effi­ciency of your work than the qual­ity. All else equal, I’m ten­ta­tively more in­ter­ested in peo­ple who can do very high-qual­ity work in­effi­ciently than peo­ple do­ing mediocre work quickly – be­cause I ex­pect that the the former are more likely to even­tu­ally do high-qual­ity, highly effi­cient work.

  • Some peo­ple tend to get very ner­vous with timed tests and mess up badly; it seems good to give them the op­por­tu­nity to prove them­selves in a less stress­ful en­vi­ron­ment.

  My cur­rent view is to ask for both timed and un­timed tests, and make the un­timed tests very sim­ple/​short (such that you could com­plete it in 20 min­utes if you had to and there’s very lit­tle benefit to spend­ing >2h on it).

• Jonas Vollmer 26 Feb 2019 19:46 UTC

  13 points

  in reply to: Khorton's comment on: After one year of ap­ply­ing for EA jobs: It is re­ally, re­ally hard to get hired by an EA organisation

  I think you’re cor­rect, but my im­pres­sion is that most EA orgs will hap­pily agree to this.

• Jonas Vollmer 26 Feb 2019 16:13 UTC

  19 points

  on: After one year of ap­ply­ing for EA jobs: It is re­ally, re­ally hard to get hired by an EA organisation

  One quick recom­men­da­tion for what ap­pli­cants can do that might be use­ful for em­ploy­ers (speak­ing from an em­ployer’s per­spec­tive): Proac­tively share pre­vi­ous work tri­als. When­ever ap­pli­cants did this, this pro­vided me with ad­di­tional valuable in­for­ma­tion that helped me de­cide whether some­one should ad­vance to the next stage.