Comments
Last week, the Forum grappled with cluelessness. The volume of entries (76 essays!)[1]—from newcomers[2] and Forum power users alike—was a happy surprise, and there were many thoughtful contributions beyond those mentioned below.
First place ($2,500):
Joint second place ($1,500 each):
Magnus Vinding (aka ‘unsharp’),[3] for Defusing cluelessness with degrees of justification
The judges—Andreas Mogensen, Jesse Clifton and Anthony DiGiovanni—decided against awarding the up-to-$50k outlier prize. Though they saw some promise in the winning essays’ approaches, their minds have not been significantly changed. The outlier prize remains open;[4] there is still much progress to be made on cluelessness.[5]
Here’s more from our judges on the winning essays. (Their comments are technical! If you haven’t yet read ‘Cluelessness: Summary of the argument’, I suggest starting there.)
This piece offers two challenges to Anthony’s argument. First, it challenges the normative premise on the grounds that only c-preferring A to B when we expect our ideal self to prefer A to B is an implausible decision rule; it can select nowhere-optimal actions, actions that aren’t optimal relative to any distribution in the representor. I’ll focus on the second challenge, which targets the inference to Anthony’s conclusion and which I found particularly interesting.
Under standard assumptions, every nowhere-optimal pure action is dominated by some mixed action: there’s some way of randomising over actions that has higher expected value according to every probability distribution in the representor. Given the author’s Unanimity and Justification principles (which seem like they ought to be plausible to the imprecise consequentialist), we are justified in c-preferring that mixture to the nowhere-optimal action. On the other hand, if an action is somewhere-uniquely-optimal – uniquely best according to some distribution in the representor – then no action is justifiably c-preferred to it. The author argues that we can sometimes identify actions of both kinds, using “flanking variants” to rule out nowhere-optimal actions and, e.g., our “forced-to-say-a-number” credences to find somewhere-uniquely-optimal candidates. This refutes the conclusion that no c-preference is ever justified and suggests a partial choice rule: exclude identifiable nowhere-optimal actions and choose an identifiable somewhere-uniquely-optimal one.
I think this is a really neat argument. Rather than arguing against imprecise consequentialism, it says that under imprecise consequentialism we can theoretically winnow out a potentially large swathe of strategies, and makes a serious attempt at turning this into a practical procedure.
One of the main reasons that this doesn’t really move me on cluelessness, yet, is that I don’t think that somewhere-uniquely-optimal actions should get much weight relative to the recommendations of other normative views. Somewhere-unique optimality is a very thin form of support: such a strategy may well look terrible according to other distributions in the representor (e.g., “keep working for a relatively responsible AI lab” might be somewhere-uniquely-optimal and yet terrible according to some other distribution, which recommends that I leave and push for an AI pause instead). I am reluctant to let this weak kind of support override other plausible normative views.
I briefly discuss two other doubts about the argument – whether we can actually distinguish nowhere-optimal and somewhere-uniquely-optimal actions, and whether imprecise consequentialists are actually required to prefer randomization – in this thread.
This essay raises a number of strong objections to the maximality rule, noting its sensitivity to outliers and small updates. The author argues that we should instead rely on a more graded approach and explores criteria based on comparing the midpoints of the distribution of expected utilities induced by the decision-maker's representor. The essay stood out for its clarity and rigour, as well as the range of ideas in play. The concerns the author notes related to the maximality rule are important and worth taking seriously, but I felt a little unsure exactly how to understand the alternative class of graded approaches, at least in the abstract. In some sense, the maximality rule is intended as answering a categorical question: should A be weakly preferred to B? On its face, that's a 'yes' or 'no' question that doesn't admit of gradation (unless perhaps we deny bivalence and introduce intermediary degrees of truth), and I’d like to have seen the author do a bit more to clarify how they see the divide here.
This piece argues that our judgements about whether one action is c-preferable to another should be graded rather than categorical. The author argues that the maximality rule, which is categorical – either one of A and B is determinately preferable to the other, or it’s indeterminate – has some implausible properties, such as an abrupt shift from determinacy to indeterminacy when a single element of the representor disagrees with the others. A graded measure, by contrast, would say that A is preferable to B to a high degree. The author then gives an example of a formula that could capture graded preference, the difference of the midpoints of the actions’ imprecise expected values, normalized by the sum of their half-widths. Finally, they say that we should wager on the strategies that do best according to such a metric, and suggest several strategies which might fit the bill: doing research to refine our estimates and expand flexible resources.
The idea that c-preferences should be graded was novel to me, and interesting. I especially appreciated how constructive the piece was: it combines a clear structural objection to maximality with explicit formal proposals and an attempt to draw practical implications.
My main concern is about action-guidingness: Although the author rightly points out that their proposal lowers the bar for action-guidance, I think that in practice the boundaries of our representor are hopelessly fuzzy, such that we won’t even be able to say whether A is determinately c-preferable to B to some degree. And although the author acknowledges the possibility of vague degrees, a main weakness of the piece, IMO, was not addressing this issue more thoroughly.
That said, this post did prompt me to consider thinking harder about whether there are cases where I would be willing to say that some rule like the ones discussed by the author gives a verdict. (As mentioned here, I think failing independence, as the author’s CD rule does, is pretty bad for something that’s supposed to track c-betterness. If I were to develop this further I’d look towards pairwise minimum regret / “midpoint-of-differences”. That rule is cyclic, but I’m not sure that (graded) c-preference need be acyclic.)
This paper defends a person-centred bracketing approach, on which one act should be preferred to another if it is better for someone and not worse for anyone, responding to concerns about this bracketing strategy raised by Kollin et al. (2025). The paper stood out for its clarity, rigour, and transparent structure, and the core points emphasized are strong and worth keeping in mind: ultimately, what we care about is the well-being of particular individuals, and not some purely abstract notion of the good. The paper is nonetheless somewhat narrow in scope. I also wasn't entirely convinced that in fact Super-Strong Pareto supports choice of A over B in the particular scenario discussed. This would seem to depend on how we individuate people. If we invent the name Twiggy for whoever would be at 20 if we choose B, then choice of A is guaranteed to be worse for Twiggy. Principles like Ex Ante Pareto and Super-Strong Pareto may thus contribute to deepening our uncertainty – this time about the proper method of individuation.
Bracketing is a theory of decision-making that makes precise this idea: we should base our decisions on consequences that we “aren’t clueless about”. This piece defends a variant of bracketing that was rejected in the original paper, bottom-up bracketing. In any variant of bracketing, we start with the individual units that standard consequentialist theories aggregate over, e.g., persons, or units of spacetime. Then, roughly, bottom-up bracketing says to ask each individual for their verdict on your choice, then go with the aggregate of the non-clueless verdicts; while top-down bracketing says to ask each possible aggregate of the individuals for its verdict, then go with the non-clueless verdict from the “largest” aggregate.
The author’s core defense of bottom-up bracketing is: Our altruistic reasons for choice fundamentally come from individual persons. So the verdicts that matter are those of the individuals. For example, suppose some person has a complaint against you choosing A, while no one else has a complaint against either option (since each of their verdicts is clueless). Why would you override that first person’s interests by choosing A, merely because the verdict of the aggregate favors A — as top-down bracketing recommends? Rather, arguably, you only have a reason to choose B, as bottom-up bracketing recommends. The author then carefully responds to several objections to (bottom-up) bracketing.
Besides its general rigor, I really appreciated this essay’s methodology. It starts by assessing views about where our reasons come from on their own merits, then unpacks what follows from there. That seems like the right approach to me, and it updates me in favor of bottom-up bracketing. That said, I think person-moments are an even more plausible candidate for the fundamental altruistic reason-givers, and would’ve liked to have seen more discussion of this.
Echoing Jesse and the author themselves, I’m unfortunately pessimistic that the most normatively well-motivated variants of bracketing (bottom-up or otherwise) will be action-guiding. For example, donating to AMF seems to be good for various far-future persons, and bad for other far-future persons, and we remain clueless about the balance of these verdicts. But gaining clarity on which variant of bracketing is normatively well-motivated in the first place seems like progress.
Beyond the winners, seven more essays made the judges’ shortlist, in rough order:
Agency and the justification of altruistic preference[6] – Marc Moffett
There are over 150 comments on the week’s posts. Prizes, as judged by Anthony, Toby Tremlett and Will Aldred (that’s me!), go to:
Honourable mentions go to Ben West for never losing sight of what could move the debate, and Jo_ for neatly summarising why cluelessness is such a problem for us EAs: “Impartial goodness is freaking weird.”[9]
If you won a prize, you don’t need to do anything. We’ll contact you about payment in the coming days.
A key hope for the competition was to surface at least one promising new researcher worth funding longer-term to work on cluelessness and/or adjacent topics. On this front, I’m cautiously optimistic.[10]
Thanks to Andreas, Jesse and Anthony for judging. Double thanks to Anthony for also writing the argument this competition invited everyone to attack. And finally, I’m grateful to all of you who took part in the discussion—30 posts and 150+ comments. That’s a lot of collective thinking on what, to my mind, is now EA’s central intellectual problem.[11]
Special shoutouts to academic philosophers Marc Moffett, Claire Field and Nicholaos Jones.
Magnus says that he published under a pseudonym because he’s interacted with each of the judges before, and blind review struck him as fairer. This is a great point, and one that totally escaped me when designing the competition.
If you develop a solution to or critique of cluelessness that you believe meets the bar of the outlier prize, email will at macroscopic dot org or DM me on this forum. I may only have time to respond to the most promising submissions. The prize will stay open until further notice, though the judging panel may change over time.
For some directions for further work, see the judges' comments on the winning essays below.
Judges noted this one as particularly well argued, but less significant or action-relevant.
Toby announced mid-week: “I’m awarding this comment from Jim [about person-moments for bottom-up bracketing] although I could have awarded many of his others – he’s bringing loads of value in his comments this week, de-confusing conversations and responding to posts with helpful clarifying questions. Thanks Jim!”
Kuutti also left informative comments on ‘The cluelessness horizon argument’ and on their own post, ‘Lexical filtering’, which narrowly missed out on an essay prize.
Original credit for this quote goes to Max Daniel.
Perhaps I’ll be able to say more in time.
Fleshing this out a little bit: i) I see Anthony’s cluelessness argument as the most crucial consideration since longtermism. I hope that last week was but a first step toward our orienting to it properly. ii) Behind cluelessness sits a broader question: what decision theory should EAs actually use, given that idealised Bayesianism doesn’t directly apply to boundedly rational agents like us? (Should we be approximating expected value? If so, what does this mean, exactly?)
