Scott Alexander

Ex­cerpts From The EA Talmud

Pause For Thought: The AI Pause Debate

• Scott Alexander 1 Sep 2023 2:35 UTC

  2 points

  0 ∶ 0

  in reply to: Omnizoid’s comment on: Eliezer Yud­kowsky Is Fre­quently, Con­fi­dently, Egre­giously Wrong

  I thought we already agreed the demon case showed that FDT wins in real life, since FDT agents will consistently end up with more utility than other agents.

  Eliezer’s argument is that you can become the kind of entity that is programmed to do X, by choosing to do X. This is in some ways a claim about demons (they are good enough to predict even the choices you made with “your free will”). But it sounds like we’re in fact positing that demons are that good—I don’t know how to explain how they have 999,999/​million success rate otherwise—so I think he is right.

  I don’t think the demon being wrong one in a million times changes much. 999,999 of the people created by the demon will be some kind of FDT decision theorist with great precommitment skills. If you’re the one who isn’t, you can observe that you’re the demon’s rare mistake and avoid cutting off your legs, but this just means you won the lottery—it’s not a generally winning strategy.

  Decision theories are intended as theories of what is rational for you to do. So it describes what choices are wise and which choices are foolish.

  I don’t understand why you think that the choices that get you more utility with no drawbacks are foolish, and the choices that cost you utility for no reason are wise.

  On the Newcomb’s Problem post, Eliezer explicitly said that he doesn’t care why other people are doing decision theory, he would like to figure out a way to get more utility. Then he did that. I think if you disagree with his goal, you should be arguing “decision theory should be about looking good, not about getting utility” (so we can all laugh at you) rather than saying “Eliezer is confidently and egregiously wrong” and hiding the fact that one of your main arguments is that he said we should try to get utility instead of failing all the time and then came up with a strategy that successfully does that.

• Scott Alexander 30 Aug 2023 3:24 UTC

  6 points

  4 ∶ 1

  in reply to: Omnizoid’s comment on: Eliezer Yud­kowsky Is Fre­quently, Con­fi­dently, Egre­giously Wrong

  I think rather than say that Eliezer is wrong about decision theory, you should say that Eliezer’s goal is to come up with a decision theory that helps him get utility, and your goal is something else, and you have both come up with very nice decision theories for achieving your goal.

  (what is your goal?)

  My opinion on your response to the demon question is “The demon would never create you in the first place, so who cares what you think?” That is, I think your formulation of the problem includes a paradox—we assume the demon is always right, but also, that you’re in a perfect position to betray it and it can’t stop you. What would actually happen is the demon would create a bunch of people with amputation fetishes, plus me and Eliezer who it knows wouldn’t betray it, and it would never put you in the position of getting to make the choice in real life (as opposed to in an FDT algorithmic way) in the first place. The reason you find the demon example more compelling than the Newcomb example is that it starts by making an assumption that undermines the whole problem—that is, that the demon has failed its omniscience check and created you who are destined to betray it. If your problem setup contains an implicit contradiction, you can prove anything.

  I don’t think this is as degenerate a case as “a demon will torture everyone who believes FDT”. If that were true, and I expected to encounter that demon, I would simply try not to believe FDT (insofar as I can voluntarily change my beliefs). While you can always be screwed over by weird demons, I think decision theory is about what to choose in cases where you have all of the available knowledge and also a choice in the matter, and I think the leg demon fits that situation.

• Scott Alexander 29 Aug 2023 18:22 UTC

  2 points

  0 ∶ 1

  in reply to: Omnizoid’s comment on: Eliezer Yud­kowsky Is Fre­quently, Con­fi­dently, Egre­giously Wrong

  I guess any omniscient demon reading this to assess my ability to precommit will have learned I can’t even precommit effectively to not having long back-and-forth discussions, let alone cutting my legs off. But I’m still interested in where you’re coming from here since I don’t think I’ve heard your exact position before.

  Have you read https://​​www.lesswrong.com/​​posts/​​6ddcsdA2c2XpNpE5x/​​newcomb-s-problem-and-regret-of-rationality ? Do you agree that this is our crux?

  Would you endorse the statement “Eliezer, using his decision theory, will usually end out with more utility than me over a long life of encountering the sorts of weird demonic situations decision theorists analyze, I just think he is less formally-rational” ?

  Or do you expect that you will, over the long run, get more utility than him?

• Scott Alexander 29 Aug 2023 5:34 UTC

  4 points

  0 ∶ 0

  in reply to: Omnizoid’s comment on: Eliezer Yud­kowsky Is Fre­quently, Con­fi­dently, Egre­giously Wrong

  Sorry if I misunderstood your point. I agree this is the strongest objection against FDT. I think there is some sense in which I can become the kind of agent who cuts off their legs (ie by choosing to cut off my legs), but I admit this is poorly specified.

  I think there’s a stronger case for, right now, having heard about FDT for the first time, deciding I will follow FDT in the future. Various gods and demons can observe this and condition on my decision, so when the actual future comes around, they will treat me as an FDT-following agent rather than a non-FDT-following agent. Even though future-created-me isn’t exactly in a position to influence the (long-since gone) demon, current me is in a position to make this decision for future relevant situations, and should decide to follow FDT in general. Part of this decision I’ve made involves being the kind of person who would take the FDT option in hypothetical scenarios.

  Then there’s the additional question of whether to defect against the demons/​gods later, and say “Haha, back in August 2023 I resolved to become an FDT agent, and I fooled you into believing me, but now that I’ve been created I’m just going to not cut off my legs after all”. I think of this as—suppose every past being created by the demon has cut off its legs, ie the demon has a 100% predictive success rate over millions of cases. So the demon would surely predict if I would do this. That means I should (now) try really hard not to do this. Cf. Parfit’s Hitchhiker. Can I bind my future self like this? I think empirically yes—I think I have enough honor that if I tell hypothetical demon gods now that I’m going to do various things, I can actually do them when the time comes. This will be “irrational” in some sense, but I’ll still end up with more utility than everyone else.

  Is there some sense in which, if I decide not to cut off my legs, I would wink out of existence? I admit feeling a superstitious temptation to believe this (a non-superstitious justification might be wondering if I’m the real me, or a version of me in the omniscient demon’s simulation to predict what I would do). I think the literal answer is no but that it’s practically useful to keep my superstitious belief in this to allow myself to do the irrational thing that gets me more utility. But this is a weird enough sidetrack that I’m really not sure I’m still in normal Eliezer-approved-decision-theory-land at all.

  I think an easier question is whether you should program an AI to always keep its pre-emptive bargains with gods and demons; here the answer is just straightforwardly yes. You don’t have to assume that your actions alter your algorithm, you can just alter the algorithm directly. I think this is what Eliezer is most interested in, though I’m not sure.

• Scott Alexander 28 Aug 2023 22:07 UTC

  6 points

  0 ∶ 0

  in reply to: Guy Raveh’s comment on: Eliezer Yud­kowsky Is Fre­quently, Con­fi­dently, Egre­giously Wrong

  Were there bright people who said they had checked his work, understood it, agreed with him, and were trying to build on it? Or just people who weren’t yet sure he was wrong?

• Scott Alexander 28 Aug 2023 22:06 UTC

  3 points

  1 ∶ 0

  in reply to: Omnizoid’s comment on: Eliezer Yud­kowsky Is Fre­quently, Con­fi­dently, Egre­giously Wrong

  I don’t want to get into a long back-and-forth here, but for the record I still think you’re misunderstanding what I flippantly described as “other Everett branches” and missing the entire motivation behind Counterfactual Mugging. It is definitely not supposed to directly make sense in the exact situation you’re in. I think this is part of why a variant of it is called “updateless”, because it makes a principled refusal to update on which world you find yourself in in order to (more flippant not-quite-right description) program the type of AIs that would weird games played against omniscient entities.

  If the demon would only create me conditional on me cutting off my legs after I existed, and it was the specific class of omniscient entity that FDT is motivated by winning games with, then I would endorse cutting off my legs in that situation.

  (as a not-exactly-right-but-maybe-helpful intuition pump, consider that if the demon isn’t omniscient—but simply reads the EA Forum—or more strictly can predict the text that will appear on the EA Forum years in the future—it would now plan to create me but not you, and I with my decision theory would be better off than you with yours. And surely omniscience is a stronger case than just reads-the-EA-Forum!)

  If this sounds completely stupid to you, and you haven’t yet read the LW posts on Counterfactual Mugging. I would recommend starting there; otherwise, consider finding a competent and motivated FDT proponent (ie not me) and trying to do some kind of double-crux or debate with them, I’d be interested in seeing the results.

• Scott Alexander 28 Aug 2023 19:50 UTC

  12 points

  5 ∶ 5

  on: Eliezer Yud­kowsky Is Fre­quently, Con­fi­dently, Egre­giously Wrong

  I won’t comment on the overall advisability of this piece, but I think you’re confused about the decision theory (I’m about ten years behind state of the art here, and only barely understood it ten years ago, so I might be wrong).

  The blackmail situation seems analogous to the Counterfactual Mugging, which was created to highlight how Eliezer’s decision theories sometimes (my flippant summary) suggest you make locally bad decisions in order to benefit versions of you in different Everett branches. Schwartz objecting “But look how locally bad this decision is!” isn’t telling Eliezer anything he doesn’t already know, and isn’t engaging with the reasoning. I think I would pay Omega in Counterfactual Mugging; I agree Schwartz’s case is harder, but provisionally I think it unintentionally adds a layer of Pascal’s Wager + torture vs. dust specks by making the numbers so extreme, which are two totally unrelated reasoning vortices.

  I think the “should you procreate to make your father procreate?” question only works if your father’s cognitive algorithms are perfectly correlated with yours, which no real father’s are. To make the example fair, it should be more like “You were created by Omega, a god who transcends time. It resolved to created you if and only if It predicted that you would procreate, and It is able to predict everything perfectly. Now should you procreate?” I would also accept “You were created by a clone of yourself in the exact same situation, down to the atom, that you find yourself in now, including worrying about being created by a clone of yourself and so on. Should you procreate?” In both of these, the question seems much more open than with a normal human father.

  If Eliezer’s decision theories make no sense and are ignoring easy disproofs, then everyone else who finds them compelling (or at least not obviously wrong) after long study, including people like Wei Dai, Abram Demski, Scott Garrabrant, Benya Fallenstein, etc, is also bizarrely and inexplicably wrong. This is starting to sound less like “Eliezer is a uniquely bad reasoner” and more like “there’s something in the water supply here that makes extremely bright people with math PhDs make simple dumb mistakes that any rando can notice.”

• Scott Alexander 23 Aug 2023 21:46 UTC

  24 points

  2 ∶ 0

  on: Select ex­am­ples of ad­verse se­lec­tion in longter­mist grantmaking

  Thanks for writing this.

  I understand why you can’t go public with applicant-related information, but is there a reason grantmakers shouldn’t have a private Slack channel where they can ask things like “Please PM me if any of you have any thoughts on John Smith, I’m evaluating a grant request for him now”?

• Scott Alexander 22 Jul 2023 19:52 UTC

  31 points

  4 ∶ 1

  on: EA Sur­vey 2022: How Peo­ple Get In­volved in EA

  Okay, so GWWC, LW, and GiveWell, what are we going to do to reverse the trend?

  Seriously, should we be thinking of this as “these sites are actually getting less effective at recruiting EAs” or as “there are so many more recruitment pipelines now that it makes sense that each one would drop in relative importance” or as “any site will naturally do better in its early years as it picks the low-hanging fruit in converting its target population, then do worse later”?

• Scott Alexander 11 Jun 2023 9:37 UTC

  33 points

  13 ∶ 0

  in reply to: Justin Sandefur’s comment on: How economists got Africa’s AIDS epi­demic wrong

  The first point seems to be saying that we should factor in the chance that a program works into cost-effectiveness analysis. Isn’t this already a part of all such analyses? If it isn’t, I’m very surprised by that and think it would be a much more important topic for an essay than anything about PEPFAR in particular.

  The second point, that people should consider whether a project is politically feasible, is well taken. It sounds like the lesson here is “if you find yourself in a situation where you have to recommend either project A or B, and both are good, but A is better than B, but if you do activism for A it still won’t happen, but if you do activism for B your input will push it over the edge into happening, do activism for B.” I agree with this as far as it goes, but there seem like some important caveats:

  • You shouldn’t lie, for the normal reasons against lying. It sounds like some of these economists were just publishing reports or articles saying that antimalarials were better than PEPFAR, and I think if that was what they found, then publishing that in reports is correct regardless of the politics. Then when they put on their activist hats they can support whatever is most effective to support.

  • It’s a really hard problem to know whether to pursue less ambitious vs. more ambitious goals. If you’re a socialist, should you spend your time fighting for a $1 minimum wage increase, for Medicare For All, or for completely dismantling capitalism and replacing it with workers’ communes? I don’t think there’s an obvious answer, even though the first is clearly more likely to succeed than the second two. In retrospect it’s clear that PEPFAR worked politically, and if you say that more cost-effective options wouldn’t have worked politically at that time then I believe you, but I don’t want to conclude that therefore less ambitious political projects are always better than more ambitious ones, and I don’t know how else to apply the lesson from this story more generally.

• Scott Alexander 6 Mar 2023 21:20 UTC

  7 points

  1 ∶ 0

  in reply to: PatrickL’s comment on: Scor­ing fore­casts from the 2016 “Ex­pert Sur­vey on Progress in AI”

  Update: I think Bing passes the high school essay bar, based on the section “B- Essays No More” at https://​​oneusefulthing.substack.com/​​p/​​i-hope-you-werent-getting-too-comfortable

• Scott Alexander 4 Mar 2023 17:10 UTC

  15 points

  1 ∶ 0

  on: Pre­dic­tive Perfor­mance on Me­tac­u­lus vs. Man­i­fold Markets

  I find this interesting, but also somewhat hard to identify any meaningful patterns. For example, one could expect red points to be clustered at the top for Manifold, indicating that more forecasts equal better performance. But we don’t see that here. The comparison may be somewhat limited anyway: In the eyes of the Metaculus community prediction, all forecasts are created equal. On Manifold, however, users can invest different amounts of money. A single user can therefore in principle have an outsized influence on the overall market price if they are willing to spend enough. I’d be interested to see more on how accuracy on Manifold changes with the number of traders and overall trading volume. Who knows, maybe Manifold would be ahead if they had a similar number of forecasters to Metaculus?

  Does this mean that if you controlled for number of forecasters, you still think Metaculus would beat Manifold? If not, do you have any opinion on this question (sorry if I missed it).

• Scott Alexander 3 Mar 2023 22:11 UTC

  8 points

  0 ∶ 0

  in reply to: PatrickL’s comment on: Scor­ing fore­casts from the 2016 “Ex­pert Sur­vey on Progress in AI”

  Thank you. I misremebered the transcription question. I now agree with all of your resolutions, with the most remaining uncertainty on translation.

• Scott Alexander 3 Mar 2023 9:38 UTC

  32 points

  1 ∶ 0

  on: Scor­ing fore­casts from the 2016 “Ex­pert Sur­vey on Progress in AI”

  Thank you for doing this! I was working on a similar project and mostly came up with the same headline finding as you: the experts seemed well-calibrated. I did decide a few of the milestones a little differently, and would like to hear why you chose the way you did so I can decide whether or not to change mine.

  • Zach Stein-Perlman from AI Impacts said that he thought “efficiently sort very large lists” and “write good Python code” were false, because the questions said it had to be done in a certain way by a certain type of neural net, and that wasn’t how it was done.

  • I was planning to count “transcribe as well as humans” as false, based on https://​​interactiveaimag.org/​​columns/​​is-ai-at-human-parity-yet-a-case-study-on-speech-recognition/​​ . Maybe the top labs could achieve this with a year of work, but I think the question specifies they need to do as well as the best human transcriptionists, and right now they don’t seem close.

  • I counted “translate as well as bilingual humans” as true based on a few quick tests of ChatGPT; I’m curious if you have some specific source for why it’s false.

  • I don’t think AI has won at Starcraft. The last word I’ve heard from this was https://​​www.extremetech.com/​​extreme/​​301325-deepminds-starcraft-ii-ai-can-now-defeat-99-8-percent-of-human-players , where AlphaStar could beat 99.8% of humans but not the absolute champions. I haven’t seen any further progress on this since 2019. Again, it’s possible that a year of concerted effort could change this, but that seems speculative. See also https://​​www.reddit.com/​​r/​​starcraft/​​comments/​​uakohx/​​why_cant_we_make_a_perfect_ai_for_starcraft/​​

  • I’m surprised you judged “high marks for a high school essay” as false; this seems like a central use case for ChatGPT and Bing/​GPT4.

  • I was planning to judge “concisely explain game play” as true, based on https://​​www.forbes.com/​​sites/​​carlieporterfield/​​2022/​​11/​​22/​​metas-ai-gamer-beat-humans-in-diplomacy-using-strategy-and-negotiation/​​, which is testing basically this skill. Also, I was able to play a partial game of chess with ChatGPT where it explained all its moves—before it started hallucinating and making moves which were impossible. Still, it seemed to have the “explanation” skill down pat! I imagine if you asked it to explain why a chess engine made a given move, it would give a pretty plausible answer.

  Beyond those quibbles—I was also looking at https://​​aiimpacts.org/​​2022-expert-survey-on-progress-in-ai/​​#Data (the dataset itself; the summary doesn’t include the milestones). This new version seems like total garbage. The experts continue to predict several of the milestones are five years out, including milestones that were achieved by ChatGPT (ie a few months after the survey) and at least one milestone that had already clearly been achieved by the time the survey was released! Unless there’s some reason to think the new crop of experts is worse than the old one, this makes me think they only did okay last time by luck/​​coincidence, and actually they have no idea what they’re doing.

  (I don’t think it works to say that the period 2017-2022.5 was predictable, but the period 2022.5-2023 wasn’t, because part of what the 2017 experts were right about was ChatGPT, which came out in late 2022.)

• Scott Alexander 24 Feb 2023 9:10 UTC

  17 points

  3 ∶ 0

  in reply to: Ofer’s comment on: Man­i­fund Im­pact Mar­ket /​ Mini-Grants Round On Forecasting

  Thanks for asking. One reason we decided to start with forecasting was because we think it has comparatively low risks compared to other fields like AI or biotech.

  If this goes well and we move on to a more generic round, we’ll include our thoughts on this, which will probably include a commitment not to oracular-fund projects that seem like they were risky when proposed, and maybe to ban some extremely risky projects from the market entirely. I realize we didn’t explicitly say that here, which is because this is a simplified test round and we think the forecasting focus makes risks pretty unlikely.

  In the unlikely event that someone proposes a forecasting project < $20,000 which we think carries significant risk, we’re prepared to take those steps this time too.

Man­i­fund Im­pact Mar­ket /​ Mini-Grants Round On Forecasting

• Scott Alexander 6 Feb 2023 6:33 UTC

  136 points

  19 ∶ 5

  on: EA, Sex­ual Harass­ment, and Abuse

  In 2018, I collected data about several types of sexual harassment on the SSC survey, which I will report here to help inform the discussion. I’m going to simplify by assuming that only cis women are victims and only cis men are perpetrators, even though that’s bad and wrong.

  Women who identified as EA were less likely report lifetime sexual harassed at work than other women, 18% vs. 20%. They were also less likely to report being sexually harassed outside of work, 57% vs. 61%.

  Men who identified as EA were less likely to admit to sexually harassing people at work (2.1% vs. 2.9%) or outside of work (16.2% vs. 16.5%)

  The sample was 270 non-EA women, 99 EA women, 4940 non-EA men, and 683 EA men. None of these results were statistically significant, although all of them trended in the direction of EAs experiencing less sexual harassment.

  This doesn’t prove that EA environments have less harassment than the average environment, since it could be that EAs are biased to have less sexual harassment for other reasons, and whatever additional harassment they get in EA isn’t enough to make up for it; the vast majority of EAs have the vast majority of interactions in non-EA environments. I tried to sort of get around this by limiting my analysis to people living in California, on the grounds that they were more likely to be plugged into EA communities and jobs. Conditional on being a woman in California, being EA did make someone more likely to experience sexual harassment, consistently, as measured in many different ways. But Californian EAs were also younger, much more bisexual, and much more polyamorous than Californian non-EAs; adjusting for sexuality and polyamory didn’t remove the gap, but age was harder to adjust for and I didn’t try. EAs who said they were working at charitable jobs that they explicitly calculated were effective had lower harassment rates than the average person, but those working at charitable jobs that they didn’t expliclitly calculate were higher. All of these subgroup analyses were very small sample size.

  Overall I am not sure that anything can be concluded from these results either way.

  I would urge everyone thinking about this question to read my original discussion of the sexual harassment survey results. It mostly focuses on professions but I think the overall conclusion is extremely relevant here too. You can also find the link to the data there in case you want to double-check my results.

  What links here?

  • Ian Turner's comment on EA’s weird­ness makes it un­usu­ally sus­cep­ti­ble to bad behavior by OutsideView (7 Feb 2023 6:42 UTC; 13 points)

Alexan­der and Yud­kowsky on AGI goals