Alex Semendinger 🔸

Tie train­ing can make DPO/​RLHF-trained AIs gen­er­al­ize better

• Alex Semendinger 🔸 29 Dec 2025 2:46 UTC

  5 points

  0 ∶ 0

  on: Where did you donate this year?

  This year I donated to (in decreasing order of amount given):

  • The EA Animal Welfare Fund

  • Alex Bores’s congressional campaign

  • Forethought

• Alex Semendinger 🔸 8 Nov 2024 12:52 UTC

  5 points

  2 ∶ 0

  on: Sig­nal­ing with Small Orange Diamonds

  Emojis in display names feels like a Twitter-native phenomenon. I think it works on Twitter because of the distinction between a @username and a Twitter handle: the latter can change frequently and is often used for jokes or puns anyway.

  So the orange diamond emoji fits in well on Twitter—even “Jeff Kaufman 🔸🏗👣🛝💡🌎”, while a little over the top, wouldn’t strike me as too unusual. But in most other settings (EA Forum, Facebook, LinkedIn, etc), where there’s less or no distinction between real names, usernames, and display names, an emoji stands out more. (Although 🔸 is visually simpler and more professional-looking than 🛝, at least.)

  A candidate rule of thumb: use the 🔸 in situations where you’re fine with people using other emojis, and don’t use it if it might start a slippery slope toward 🔸🏗👣🛝💡🌎 where that would be unwelcome. For me that means … just Twitter, I think? And maybe the EA forum where it’s already catching on and doesn’t seem to be spurring other emoji-use.

• Alex Semendinger 🔸 21 Mar 2024 18:23 UTC

  1 point

  0 ∶ 0

  in reply to: Sam_Coggins’s comment on: Cri­tique of the no­tion that im­pact fol­lows a power-law distribution

  Unless I’m misunderstanding, isn’t this “just” an issue of computing Shapley values incorrectly? If kindling is important to the fire, it should be included in the calculation; if your modeling neglects to consider it, then the problem is with the modeling and not with the Shapley algorithm per se.

  Of course, I say “just” in quotes because actually computing real Shapley values that take everything into account is completely intractable. (I think this is your main point here, in which case I mostly agree. Shapley values will almost always be pretty made-up in the best of circumstances, so they should be taken lightly.)

  I still find the concept of Shapley values useful in addressing this part of the OP:

  Impact does not seem to be a property that can sensibly be assigned to an individual. If an individual (or organisation) takes an action, there a number of reasons why I think that the subsequent consequences/​impact can’t solely be attributed to that one individual.

  I read this as sort of conflating the claims that “impact can’t be solely attributed to one person” and “impact can’t be sensibly assigned to one person.” Shapley values help with assigning values to individuals even when they’re not solely responsible for outcomes, so it helps pull these apart conceptually.

  Much more fuzzily, my experience of learning about Shapley values took me from thinking “impact attribution is basically impossible” (as in the quote above) to “huh, if you add a bit more complexity you can get something decent out.” My takeaway is to be less easily convinced that problems of this type are fundamentally intractable.

• Alex Semendinger 🔸 15 Mar 2024 17:31 UTC

  2 points

  0 ∶ 0

  in reply to: Sarah Weiler’s comment on: Cri­tique of the no­tion that im­pact fol­lows a power-law distribution

  I agree with just about everything in this comment :)

  (Also re: Shapley values—I don’t actually have strong takes on these and you shouldn’t take this as a strong endorsement of them. I haven’t engaged with them beyond reading the post I linked. But they’re a way to get some handle on cases where many people contribute to an outcome, which addresses one of the points in your post.)

• Alex Semendinger 🔸 15 Mar 2024 13:25 UTC

  18 points

  5 ∶ 0

  on: Cri­tique of the no­tion that im­pact fol­lows a power-law distribution

  Thanks for writing this! “EA is too focused on individual impact” is a common critique, but most versions of it fall flat for me. This is a very clear, thorough case for it, probably the best version of the argument I’ve read.

  I agree most strongly with the dangers of internalizing the “heavy-tailed impact” perspective in the wrong way, e.g. thinking “the top people have the most impact → I’m not sure I’m one of the top people → I won’t have any meaningful impact → I might as well give up.” (To be clear, steps 2, 3, and 4 are all errors: if there’s a decent chance you’re one of the top, that’s still potentially worth going for. And even if not—most people aren’t—that doesn’t mean your impact is negligible, and certainly doesn’t mean doing nothing is better!)

  I mostly disagree with the post though, for some of the same reasons as other commenters. The empirical case for heavy-tailed impact is persuasive to me, and while measuring impact reliably seems practically very hard /​ intractable in most cases, I don’t think it’s in principle impossible (e.g. counterfactual reasoning and Shapley values).

  I’m also wary of arguments that have the form “even if X is true, believing /​ saying it has bad consequences, so we shouldn’t believe /​ say X.” I think there are usually ways to incorporate X while mitigating the downsides that might be associated with believing it; some of the links you included (e.g. Virtues for Real-World Utilitarians, Naive vs Prudent Utilitarianism) provide examples of this. Heavy-tailed impact is (if true) a very important fact about the world. So it’s worth putting in the effort to incorporate it into our beliefs effectively, doing our best to avoid the downsides you point out.

  What links here?

  • Cri­tique of the no­tion that im­pact fol­lows a power-law distribution by Sarah Weiler (14 Mar 2024 10:28 UTC; 89 points)

• Alex Semendinger 🔸 9 Feb 2024 1:22 UTC

  7 points

  2 ∶ 0

  in reply to: CAISID’s comment on: Quick PSA: Edit­ing and Re­view peer group for fo­rum posts

  The FTX collapse took place in November 2022. Among other things, this resulted in a lot of negative media attention on EA.

  It’s also worth noting that this immediately followed a large (very positive, on the whole) media campaign around Will MacAskill’s book What We Owe the Future in summer 2022, which I imagine caused much of the growth earlier that year.

• Alex Semendinger 🔸 7 Feb 2024 17:25 UTC

  6 points

  0 ∶ 0

  on: Look­ing for: songs or art that in­spire EA values

  Many of the songs associated with Secular Solstice^[1] have strong EA themes, or were explicitly written with EA in mind.
  
  A few of the more directly EA songs that I like:

  • “Brighter Than Today,” the prototypical Solstice song, about the possibility of a brighter future.

  • “The Circle,” about the expanding moral circle (more literal /​ less subtle than “Draw the Circle Wide,” which I also like!)

  • “Five Thousand Years” and “Here and Now” for longtermism.

  • “Hymn to Breaking Strain” (a setting of a Kipling poem) is not about doing good per se, but feels extremely EA to me.

  1. ^

     Lots of resources at that link, also an overlapping list of solstice songs here.

• Alex Semendinger 🔸 25 Sep 2023 13:59 UTC

  13 points

  0 ∶ 0

  on: How have you be­come more hard-work­ing?

  Setting Beeminder goals for the number of hours worked on different projects has substantially increased my productivity over the past few months.

  I’m very deadline-motivated: if a deadline is coming up, I can easily put in 10 hours of work in a day. But without any hard deadlines, it can take active willpower to work for more than 3 or 4 hours. Beeminder gives me deadlines almost every day, so it takes much less willpower now to have productive days.

  (I’m working on a blog post about this currently, which I expect to have out in about two weeks. If I remember, I’ll add a link back to this comment once it’s out.)

  Update: This strategy only worked for a few months before I got pretty burnt out; having non-negotiable deadlines every day eventually became more draining than motivating. Beeminder has still helped me a ton overall, but it’s very important to get the goals right and tweak them if they stop working!

• Alex Semendinger 🔸 3 Feb 2023 20:19 UTC

  11 points

  4 ∶ 0

  on: Puz­zles for Some

  Interesting post! But I’m not convinced.

  I’ll stick to addressing the decision theory section; I haven’t thought as much about the population ethics but probably have broadly similar objections there.

  (1) What makes STOCHASTIC better than the strategy “take exactly N tickets and then stop”?

  • Both avoid near-certain death (good!)

  • Both involve, at some point, turning down what looks like a strictly better option

    • To me, STOCHASTIC seems to do this at the very first round, and all subsequent rounds. (If I played STOCHASTIC and drew a non-black ball first, I think I’d be pretty disappointed. This indicates to me that I didn’t actually want to randomize on that round.) Or you could just view STOCHASTIC as doing this on the round when you stop accepting tickets.

    • This very fact makes these strategies about as puzzling as the “reckless” or “timid” strategies to me—at some point, you’re deliberately choosing the worse option, by your own lights! That’s at least weird, right?

  • “Take exactly N” has the advantage of letting you decide the exact level of risk you want to take on, while STOCHASTIC involves an additional layer of uncertainty, which gets you ….???^[1]

  I get that you’re trying to avoid totalizing theoretical frameworks, but you also seem to be saying it’s better in some way that makes it worth choosing, at least for you. But why?

  (2) In response to

  But, well, you don’t have to interpret my actions as expressing attitudes towards expected payoffs. I mean this literally. You can just … not do that.

  I’m having trouble interpreting this more charitably than “when given a choice, you can just … choose the option with the worse payoff.” Sure, you can do that. But surely you’d prefer not to? Especially if by “actions” here, we’re not actually referring to what you literally do in your day-to-day life, but a strategy you endorse in a thought-experiment decision problem. You’re writing as if this is a heavy theoretical assumption, but I’m not sure it’s saying anything more than “you prefer to do things that you prefer.”

  (3) In addition to not finding your solution to the puzzle satisfactory,^[2] I’m not convinced by your claim that this isn’t a puzzle for many other people:

  Either you’re genuinely happy with recklessness (or timidity), or else you have antecedent commitments to the methodology of decision theory — such as, for example, a commitment to viewing every action you take as expressing your attitude to expected consequences.

  To me, the point of the thought experiment is that roughly nobody is genuinely happy with extreme recklessness or timidity.^[3] And as I laid out above, I’d gloss “commitment to viewing every action you take as expressing your attitude to expected consequences” here as “commitment to viewing proposed solutions to decision-theory thought experiments as expressing ideas about what decisions are good” — which I take to be nearly a tautology.

  So I’m still having trouble imagining anyone the puzzles aren’t supposed to apply to.

  1. ^

     The only case I can make for STOCHASTIC is if you can’t pre-commit to stopping at the N-th ticket, but can pre-commit to STOCHASTIC for some reason. But now we’re adding extra gerrymandered premises to the problem; it feels like we’ve gone astray.

  2. ^

     Although if you just intend for this to be solely your solution, and make no claims that it’s better for anyone else, or better in any objective sense then … ok?

  3. ^

     This is precisely why it’s a puzzle—there’s no outcome (always refuse, always take, take N, stochastic) that I can see any consistent justification for.

  What links here?

  • Richard Y Chappell🔸's comment on Puz­zles for Some by Violet Hour (3 Feb 2023 20:43 UTC; 12 points)

• Alex Semendinger 🔸 27 Oct 2022 23:34 UTC

  23 points

  1 ∶ 0

  on: Pod­cast: The Left and Effec­tive Altru­ism with Habiba Islam

  Another podcast episode on a similar topic came out yesterday, from Rabbithole Investigations (hosted by former Current Affairs podcasts hosts Pete Davis, Sparky Abraham, and Dan Thorn). They had Joshua Kissel on to talk about the premises of EA and his paper “Effective Altruism and Anti-Capitalism: An Attempt at Reconciliation.”

  This is the first interview (and second episode) in a new series dedicated to the question “Is EA Right?”. The premise of the show is that the hosts are interested laypeople who interview many guests with different perspectives, in the hopes of answering their question by the end of the series.

  I’m optimistic about this podcast as another productive bridge between the EA and lefty worlds; their intro episode gave me a lot of hope that they’re approaching this with real curiosity.

  (I’m posting this more to recommend /​ mention the series rather than the particular episode though; the episode itself spent most of its runtime covering intro-EA topics in what I felt was a pretty standard way, which most people here probably don’t need. That said, it did a good job of what it was aiming for, and I’m looking forward to the real heavy-hitting critiques and responses as the series continues.)

• Alex Semendinger 🔸 7 Oct 2022 2:17 UTC

  7 points

  0 ∶ 0

  on: [Creative Non­fic­tion] The Toba Su­per­vol­canic Eruption

  I read this piece a few months ago and then forgot what it was called (and where it had been posted). Very glad to have found it again after a few previous unsuccessful search attempts.

  I think all the time about that weary, determined, unlucky early human trying to survive, and the flickering cities in the background. When I spend too long with tricky philosophy questions, impossibility theorems, and trains to crazytown, it’s helpful to have an image like this to come back to. I’m glad that guy made it. Hopefully we will too!

• Alex Semendinger 🔸 6 Aug 2022 1:47 UTC

  2 points

  0 ∶ 0

  on: Is emo­tional-based giv­ing against EA?

  An important principle of EA is trying to maximize how much good you do, when you’re trying to do good. So EAs probably won’t advise you to base most of your charitable giving on emotional connection (which is unlikely to be highly correlated with cost-effectiveness) -- instead, according to EA, you should base this on some kind of cost-effectiveness calculation.

  However, many EAs do give some amount to causes they personally identify with, even if they set aside most of their donations for more cost-effective causes. (People often talk about “warm fuzzies” in this context, i.e. donations that give you a warm fuzzy feeling.) In that sense, some amount of emotion-based giving is completely compatible with EA.

• Alex Semendinger 🔸 14 May 2022 12:07 UTC

  24 points

  0 ∶ 0

  on: If EA is no longer fund­ing con­strained, why should *I* give?

  There have been a few posts discussing the value of small donations over the past year, notably:

  1. Benjamin Todd on “Despite billions of extra funding, small donors can still have a significant impact”

  2. a counterpoint, AppliedDivinityStudies on “A Red-Team Against the Impact of Small Donations”

  3. a counter-counterpoint, Michael Townsend on “The value of small donations from a longtermist perspective”

  There’s a lot of discussion here (especially if you go through the comments of each piece), and so plenty of room to come to different conclusions.

  Here’s roughly where I come out of this:

  • What’s the relevant counterfactual? Many of these comment threads turn into discussions about earning-to-give vs direct work, but if you have $1000 in your hand, ready to donate, that’s not the relevant question. Rather, you should ask, “if I don’t donate this, what would I do with it instead, and how much impact would that have?”

  • You say “I know that professional grant makers think that last-dollar funding is not cost effective because they aren’t funding more projects, but aren’t out of dollars.” I think this frames the issue incorrectly. It’s not that big funders know that other projects aren’t cost-effective, it’s that they don’t currently have enough projects that clear a certain cost-effectiveness bar. But crucially, that bar is still far above zero!

  This means

  • there are probably many opportunities that are just as cost-effective that they haven’t found (potentially you have information they don’t that you could exploit; see this section of the above ADS post)

  • marginal donations should have a cost-effectiveness at worst just below that bar, which means you’re only doing a little worse than the big funders. (This point taken from Benjamin Todd here.)