Dawn Drescher

• Dawn Drescher 9 Jul 2026 12:47 UTC

  3 points

  0 ∶ 0

  on: What’s wrong with grant­mak­ing?

  Usually, the people who are best positioned to do this are high-context in AI safety, and have been thinking about what’s missing in the field for a while.

  Hi, that’s me! I’ve applied as a Field Strategist. It’s refreshing to see people think along the same lines! (For the second time really: same with Hypercerts/​Impact Markets.)

  It’s been exasperating to go 5–8 years pitching various ideas to people and hoping someone will fill various gaps in the ecosystem because I’m too busy with my own projects to do it myself. I really want to go meta and facilitate this incubation process more intentionally.

• Dawn Drescher 23 Jun 2026 21:49 UTC

  5 points

  0 ∶ 0

  on: New Video from AI in Con­text: The Fall and Rise of Sam Altman

  I once tried to oust someone who’s like a mini version of Sam Altman and lost too. Made me feel a lot of kinship with Helen Toner when this happened.

• Dawn Drescher 22 Jun 2026 11:36 UTC

  7 points

  0 ∶ 0

  on: A note to fu­ture effec­tive char­ity founders: con­sider za­kat com­pli­ance from the start

  Interesting! It doesn’t seem too costly to implement these requirements.

  I had to google Zakat though:

  Zakat is a mandatory Islamic duty of almsgiving. Every sane, adult Muslim whose wealth exceeds a minimum threshold (called the Nisab) must donate a specific portion of their accumulated assets—typically 2.5%—to help the poor and needy. It is the third of the Five Pillars of Islam.

• Dawn Drescher 21 Jun 2026 18:24 UTC

  4 points

  0 ∶ 0

  in reply to: Ozzie Gooen’s comment on: QURI is mov­ing into main­te­nance mode

  Surplus sounds useful!

  I think everything hinges on the funding unfortunately…

  Most of the projects on my list require some $200–500k in the first year to get started, and then can scale to a few million per year over time. The large-scale retrofunding needs to start higher – $10m might work, $100m works for the XPrize, $1b could be the goal.

  The natural starting point is the incubator itself, which falls into the $200–500k range, but more towards the upper end to provide seed funding for the incubated projects.

  Why did Guesstimate/​Squiggle as for-profit not work out?

  I’d love to have a call and catch up in any case! I’m curious whether you already have an opinion on whether places like DeepMind will be interested in paying for evals like the two types mentioned here (character and backdoors).

• Dawn Drescher 20 Jun 2026 10:24 UTC

  11 points

  0 ∶ 2

  on: QURI is mov­ing into main­te­nance mode

  I’d like to throw my hat in the ring and indicate that I’d at least find it very interesting to take over for you to ensure that QURI’s mission continues! I’m currently trying to get back into the AI safety grantmaking space, but that’ll most likely fail, in which case I would welcome a plan B.

  I imagine that the grantmaking bottleneck is overblown – that a 100–1000x increase in grantmaking capacity is easily achievable through hiring obvious candidates (10x) and streamlining the processes through retroactive funding (10–100x). If the funds actually end up doing that, it’ll be better again to contribute as a charity entrepreneur, and the plan B would become a plan A.

  I’d prefer to expand QURI to projects that have less to do with quantification and more with ranking and clustering, and to adopt more of an incubator-like approach where successful projects turn into spin-offs with their own legal structure over time to introduce more resilience through redundancy. (And more Python.)

  That’ll probably require about $1m in funding over ~2 years. Is that realistic? Also I’m not sure if it clashes with your vision?

• Dawn Drescher 19 Jun 2026 13:46 UTC

  2 points

  0 ∶ 0

  in reply to: Miles Tidmarsh’s comment on: Com­mu­nity Polls on Align­ment Controversies

  Thanks! Then I don’t think I need to update my answers. I’m looking forward to your next batch of questions!

• Dawn Drescher 18 Jun 2026 16:26 UTC

  7 points

  0 ∶ 0

  on: Com­mu­nity Polls on Align­ment Controversies

  Thanks for surveying this! <3

  1. I feel like people use “AI alignment” very different. When I talk to the types who are interested in decision theory and agent foundations, they usually have something really sophisticated in mind with AIs that somehow (no known solution because I’m not happy with any of the implementations of UDT that I’ve seen) try to act in such a way as to actually produce evidence that what they want to maximize will be maximized. Other people usually just mean something like “The AI tries to act sort of like a well-intentioned person would.” The first seems good but very very hard; the second seems outright dangerous, depending on details such as the particular idealizations that are applied.

  2. Hence questions like “AI alignment to humans will in practice avoid moral catastrophes …” is a strong no for me because it might not only not prevent but actually produce those catastrophes in the first place.

  3. Idealizations to eliminate the scope insensitivity bias and idealization to eliminate the speciesist and substratist biases are two different kinds of idealizations. My answer changes radically depending on whether they can be disentangled.

  4. Regarding tractability of digital minds work – I’m unsure whether I should count my worries about backfire risks as something that reduces scope or something that reduces tractability.

  5. Regarding the reflective equilibrium, it’s critical to me whether we artificially study the TAI in isolation, which won’t happen in practice, or whether we embed it with other, different agents. The first is probably meant; the second is more pragmatic.

  6. Control strikes me as safer, easier, and less reliable – a stopgap that can buy us a few years. I like that a lot more than an incomplete alignment solution that can backfire.

  7. Suffering risks – vastly more likely in the multipolar world we’re steering towards – strike me as vastly worse than just competing away > 90% of net value, so my max. agree vote feels like an understatement. On the other hand, “will” is a higher probability than what I assign to s-risk (“might”).

Tac­ti­cal and Oper­a­tional Ex­plo­ra­tory Model­ing for AI Governance

On Re­spon­si­bil­ity and Death: Can We See Real­ity for What It Is or Will It Break Us

• Dawn Drescher 10 Jun 2026 0:27 UTC

  2 points

  1 ∶ 0

  on: Stop Ap­ply­ing And Get To Work

  I’m totally aligned in spirit and have a track record of starting project after project because no one else does.

  But I find that for most of the projects that I wish existed, you either need money or fame to get them started. I don’t mean a bit of money for the rent but like $1m to incentivize downstream projects or the kind of fame that allows you to beat the cold start problem because it lets you motivate enough people to all try something new.

  I’m putting my own projects on the backburner now to focus on applications that can put me in a position where I’m better able to fund them.

• Dawn Drescher 22 May 2026 14:01 UTC

  4 points

  0 ∶ 0

  in reply to: Mo Putera’s comment on: De­nis Drescher’s Shortform

  Ohhh! I love that post! Gotta link it to my excessively guilty friends! <3

• Dawn Drescher 22 May 2026 11:32 UTC

  35 points

  1 ∶ 0

  on: De­nis Drescher’s Shortform

  Over the past 12 years, I almost always avoided applying for any jobs in effective altruism – though they did often seem like dream jobs – because:

  1. I was afraid I might not be the best candidate, and if, by chance, I replace the best candidate, my work would not only be a waste but outright harmful. I’m the sort of person who’s afraid to drive a car for fear of hurting someone, and the funding allocation can affect the lives of billions or trillions of beings, so any mistakes I could make would vastly outstrip any harm I could do with a car if I tried.

  2. That the best candidate might not make up for that harm in some other job that they do instead because they might be more socially motivated than me and not fall back on earning to give if they don’t find a charity job but rather value drift and do some mainstream stuff – in the worst case AI capabilities.

  3. That I can survive many years of earning to give without value-drifting because I have managed to do that in the past (a USP I should capitalize on because the counterfactual of the money that I earn at a random company is very low impact, so I can generate great counterfactual impact rather than a bit of marginal impact that I’d get at a charity).

  4. That applying for a job, being considered the best candidate, and then not living up to the expectations would feel deeply humiliating – I’d feel like fraud, feel guilty for the harm I’ve caused, feel ashamed of having betrayed all the people at the organization, feel like I can never live it down or risk running into any of them ever again at conferences and such.

  5. That sometimes they end up hiring a world-renowned researcher like Carl Shulman, and then I’d feel ashamed of even having considered applying because just the thought of it already feels hubristic.

  The upshot for me was:

  1. To apply to places that have the funding and management capacity to hire everyone above some bar.

  2. To time the boom and bust cycles in EA and go into grantmaking at the start of the boom cycles and ETG during the bust cycles.

  3. To trade off funding vs. management capacity, and try to contribute to grantmaking in a way that doesn’t come at a cost in management capacity, e.g., not as employee, at times when management capacity is more limiting than funding. Or to create management capacity.

  4. To wait for someone to reach out to me unprompted to apply for a role, because then they’ve already decided that I might be a good fit and thus taken some of the terrible responsibility off my shoulders.

  5. A friend of mine does a lot of drunk driving and usually goes far above the speed limit. I sometimes meet with friends for optional pastime activities. The risk that I catch a cold and my performance is degraded after such a hangout is vastly more severe than the risk of my friend’s drunk speeding because of all the hundreds of thousands of lives it might affect. Conversely, my friend hasn’t killed a single person yet. So I’m in no position to judge her.

  Meanwhile rejections were not a problem for me, so it’s not really “rejection sensitivity.” I talked to my friends about how I’m expected to react to them, and their advice was helpful. If I had dared to apply for particularly responsible roles, a rejection would’ve been a relief. After all, rejection is a return to safety. It’s just mixed with the shame over whatever mistakes I must’ve made in front of the interviewers. I considered not going to any conferences anymore where I might run into them, but my friends told me that’s unnecessary. And it’s true because when I decided not to hire people at my companies, I didn’t want them to avoid me afterwards even if they’ve made mistakes in the interviews.

  But more recently I’ve updated in the following ways:

  1. I heard from some hiring managers I’m friends with that even at their EA charities some applicants lie about their qualifications. It’s like these people want to cause mayhem and destruction on a global scale by replacing better candidates. Naturally, there friends didn’t hire the liars, but who knows how many slipped through and didn’t get caught. I imagine these people are rare, so the probability that I’ll replace someone better than me is higher but the severity of getting replaced by a liar is worse. So there’s a tradeoff I didn’t consider.

  2. It seems erroneous to think that someone can likely be a better candidate than me and yet value drift more easily than me. It’s not impossible, but it’s a strange convergence of mildly contradictory traits that I should thus have discounted. Besides, small-scale 1:1 compassion has a strong pull for me, so I run a risk of value-drifting away from high scalability impact to something like therapy.

  3. If I implicitly compare myself to Carl Shulman’s polished outputs from the past decades, I set an unrealistically high bar and will necessarily feel lacking. It stands to reason that the actual best candidate will also make mistakes and that even Carl Shulman has made mistakes. Plus, there are not enough Carl Shulmans for every position at every organization.

  4. I’ve practiced imagining that somehow I’ve ended up committing horrendous war crimes on a global scale and how I would process the guilt and try to make it up to all the families I’ve destroyed. It helps me make my fears concrete like that and think through, step by step, how I would manage such a situation responsibly.

  5. If you’re smart enough, self-deceptions won’t be obviously false, sometimes not false at all, but they’ll be selective and suspiciously purposeful.

  6. I have some information that HR doesn’t have but HR also has some information that I don’t have. Ideally, we collaborate to make the optimal decision.

  Finally, for anyone struggling with similar difficulties in the face of overwhelming responsibility, here’s a small example of someone processing his responsibility for a terrible accident.

• Dawn Drescher 18 May 2026 17:58 UTC

  6 points

  2 ∶ 3

  in reply to: Brad West🔸’s comment on: New EA Fo­rum LLM-use policy

  Exactly. I care that the ideas in the post are good. Who does the actual “typing” is irrelevant for me.

  And ghostwriting is nothing new. It’s somewhat common practice for busy top researchers to verbally discuss topics with more junior researchers who still have a good grasp on what they’re talking about and then for the junior researchers to write there verbal discussions down in the form of an easily digestable article.

  Perhaps there are reasons why ghostwriting should be disclosed or not, but I don’t see why ghostwriting by an AI is a special case that deserves more attention than human ghostwriting.

• Dawn Drescher 11 May 2026 9:52 UTC

  3 points

  0 ∶ 0

  in reply to: Jo_🔸’s comment on: Reflec­tions on An­thropic and EA

  Found it! So cool! 🦋💃

• Dawn Drescher 10 May 2026 21:30 UTC

  3 points

  1 ∶ 0

  on: Reflec­tions on An­thropic and EA

  Right now, I’m leaving my job to go start a new charity, working on a very narrow intervention.

  What is it? I’m curious!

• Dawn Drescher 7 May 2026 23:21 UTC

  2 points

  0 ∶ 0

  in reply to: Tristan Katz’s comment on: Welfare Biol­ogy and AI: Soil and Sea

  Oof! That’s a big variation! Thanks for flagging that!

• Dawn Drescher 7 May 2026 23:10 UTC

  2 points

  0 ∶ 0

  in reply to: Hans Schönberg 🔸’s comment on: Welfare Biol­ogy and AI: The Psy­chopath, the Ne­ma­tode, and the Arahant

  The purpose of the whole first article is to lay out all these perspectives and make mine transparent. Originally I wanted to just write the sequence, but Claude suggested to add a whole ’nother post in front to introduce people to the ethical foundations.

  We’ve had the discussion of ethical systems on this article. One of my core values is cooperativeness, and while I intuitively started out as classic utilitarian, I found it at odds with many common and widely held value systems and not in ways where I would’ve been ready to bite the bullet. E.g., it maps to Quiverfull, which is a rather extreme outlier. I also don’t bite the bullet on the Repugnant Conclusion.

  But most people don’t even think in terms of unbounded maximization or minimization, and certainly not in terms of classic utilitarianism. It’s such a niche view out there. Even if I talk about implications that I actually like, like tiling the universe with blazingly fast, happy computations, people look at me weird. If I want to engage in moral trades with them, I have to adjust to a wholly different viewpoint. So it was probably around 2013–14 or so that I felt compelled to abandon classic utilitarianism as extreme, niche, and untenable.

  I never settled on anything new after that but rather tried to figure out what the smallest common denominator is of the largest possible subset of all ethical preferences that I could think of.

  The smallest common demoninator that is widely shared is that suffering is bad. Lots of people care about lots of other stuff on top, and sometimes it’s not the most important thing for them, but most people seem to be in agreement that suffering is bad. That’s also why I got so interested in sadism, an important exception from the rule.

  So I think by focusing just on “suffering is bad, let’s reduce it” (and the particular framing around preferences that I find more intuitive), I can make my articles relevant to the widest audience possible. The Quiverfulls will have quibbles with it, the sadists will hate it, but maybe some 90+% of the world will benefit from them.

Welfare Biol­ogy and AI: The AI Eats the Sun

• Dawn Drescher 5 May 2026 15:02 UTC

  2 points

  0 ∶ 0

  in reply to: Hans Schönberg 🔸’s comment on: Welfare Biol­ogy and AI: What We Can Do Now

  You adapt them to the other species. We can do the same with much better information with our friends: One friend enjoys tomato sauce; another friend doesn’t enjoy it. You want to get pizza for all three of you, and it’s supposed to be a surprise, so instead of being like, “Pizza has tomato sauce on it, and my friend doesn’t like it, so I can’t make any inference about whether they want pizza and should [get pizza for them anyway, get pizza only for myself and the other friend],” you can make adjustments like, “Pizza has tomato sauce on it, and my friend doesn’t like it, so let’s ask whether they can make pizza without tomato sauce.”

  E.g., I give this example of eusocial vs. solitary insects. So when I wonder whether it’s stressful for an insect to be caught in my flat and not finding the way out (despite some nutritious stuff I have lying around), my guess is that it’s more likely stressful for the eusocial one who wants to get back to their hive than for the solitary one. When I had a fly over for a week or so in 2021, I gave her a name, enjoyed her company, and didn’t worry much about her feeling trapped (I did leave a window open when I was awake), but when I had a bee over a few days ago, I was more concerned and helped the bee find their way out.

  Likewise, humans can have some 0–3 children, cows some 5–10, and cats some 60–100 over their lifetimes. They all parent and alloparent. Turtles more like 1–2k, and no (allo-) parenting. So (just based on this data) I imagine that the loss of a child is almost as bad for a cow as it is for a human, that cats grieve somewhat less, and turtles not at all.

  Or when it comes to pain, I mention in my last article that it probably doesn’t make sense to have very strong pain signals, when an animal cannot react to them. So for nematodes it’s not very useful to experience pain strongly; for a fly it may be very useful.

  In your thought experiment there is a disconnect there that is crucial for my work: Antifrustrationism, as a form of preference utilitarianism, is all about the preferences of the individuals. If someone can fulfill their own preferences, there’s nothing for me to do. If someone can tell me how I can help them fulfill their preferences, I’ll happily do things for them that strike me as odd, like unusual kinks, because even though I don’t share them, I trust their ability to know and communicate their preferences.

  I only run into problems in cases where there is a power differential but the beings cannot communicate their preferences, e.g., because they are in the future, far away in the universe, or it’s hard/​impossible to build the sorts of experiments where one could elicit revealed preferences. Those are the cases where it gets uncomfortable because I need to make guesses and inferences.

  It’s like with a caring mother of an infant who can’t talk yet: The infant can’t help themselves (yet), so she has to do it, but to do that well, she has to infer the preferences rather than asking what they are.

  With the aliens, the situation is reversed. They are powerful but really dumb and could just ask how they can really help us, and NPP is not an obvious lever for a K-strategist species. But we have this problem with aging. Maybe they’ll stop by Earth 100k years ago, largely fail to communicate with us, but study our biology and notice that this aging thing really sucks, especially in the last couple of decades of an animal’s life. So just leave a time capsule with information on how to reverse aging that’s designed in such a way that they hope we can decypher it once we have the necessary foundational technologies.

• Dawn Drescher 5 May 2026 13:55 UTC

  2 points

  0 ∶ 0

  in reply to: Jonas Hallgren 🔸’s comment on: Notes on equa­nim­ity from the inside

  A confusing thing about my experience is that the truth of the no-self state is intellectually ineluctable to me and yet my perception is still filtered by selfhood. Sometimes I get a burst of spite-fueled energy that almost collapses into fatalism, and I remind myself that I chose this, that I can just go back if I don’t like it, but that I want to keep reading this chapter of my life, as it were, because it’s exciting! But even when I introspect, there’s a perspective there – someone has called it the watcher?