Thatâs a fair challenge.
It feels hard to come up with simple neat examples of clueless effects, which is why I left it ambiguous in my original question. But I think we can make it more precise without shifting the intuition (or at least my intuition).
Letâs say effect 3 is that your act will shift the probability of dystopian lock in (a many thousand year dictatorship) in a direction that you are clueless about.
My intuition still feels exactly the same.
Thanks for the reply!
I donât think I agree that her welfare based answer is obviously ânoâ. If I put myself in her shoes, I think I would choose not to be cut by the glass.
I am considering whether something should happen that will:
Cause me a nasty injury with broken glass.
Have some other larger impact on my welfare that I am clueless about the sign of.
Why is it incoherent for me to say:
âThe part of the effect that I am not clueless about is very bad, so I prefer that it does not happen.â
..?
It is not clear to me why this is wrong. I admit it is also not clear that this reasoning is coherent. This is a confusing thing to reason about. But I think claiming I can have no welfare based reason for preferring one outcome here needs more justification, it should not just be asserted (is this not ultimately the kind of question that this whole essay competition is about?)
If I was in this situation, I think I would choose not to be cut by the glass, for welfare based reasons. If this choice meant that someone else missed out on a tasty snack, I would not mind a theory that told me it was indeterminate whether my choice was still the best one. But if a theory tells me that giving the other person the snack is the strictly better choice, then this feels wrong.
Interesting!
I think we can just add that I am confident effect 3 will not become relevant until after I am dead. I didnât explicitly state that assumption, but that was what I had in mind. Then we get the problematic conclusion I was worried about I think?
This also raises another interesting question though, which is how bracketing interacts with personal identity being fuzzy. If bracketing cares about aggregating the effects on one âpersonâ then how do we deal with situations involving branching personal identity? I would have thought you would have to frame it in terms of âperson momentsâ, rather than âpersonsâ, in order to be coherent?
Thereâs something that bothers me with bracketing, that Iâd be interested to hear a reply to. It seems to affect bottom up bracketing as you have presented it here. I am unsure if it affects top down bracketing as well, but it sounds like it might.
Suppose I have an action available to me that will have the following 3 effects, impossible to disentangle:
I get a tasty snack right away.
Some broken glass is left somewhere. I am confident it will remain buried for 100 years, but after that I think it is highly likely to resurface and cause someone a nasty injury, due to where it is located.
Some effect on the long term future which I think will be significant in expectation, but which I am clueless about the sign of.
My understanding is that bracketing would say not only is this act permissable, but that I should do it. This is because it benefits me, and I am clueless about its effects on every other beneficiary (the person who is injured by the glass in 100 years may well be compensated by a benefit from 3).
But this feels wrong! If I just consider the effects that I am not clueless about, the bad seems to clearly outweigh the good. The same intuition that motivated bracketing in the first place would seem to say that this act is wrong.
I can buy the plausibility of a theory that says it is indeterminate whether I should take this act or not, due to the uncertainty bundled up in 3. And I can buy a theory which says the act is wrong (since that is what my intuition strongly points towards). But a theory that says it is wrong not to take this act feels like it has gone wrong somewhere.
Apologies if this is naive and has already been addressed at length somewhere. I would be interested to know how proponents of bracketing respond to this.
I see, thanks!
I found this summary really helpful, thank you!
One thing Iâm a bit confused by: in the table you classify several objections to incomparability as being responses to premise 1. But when I read premise 1 it doesnât obviously seem to make any claim about incomparability? That seems to enter in premise 2. Am I missing something?
Really interesting write up, thanks! What would the effects be on aviation?
As a data scientist who frequently wonders about this question, thank you for this post, it was a really interesting read!
But Iâm a bit confused about your final section on âgenerative AIâ roles, and the claim that data people have no special advantage here. What sort of roles do you have in mind exactly? If youâre talking only about prompt engineering, or just about the skill of being able to use a product like Claude code/âcowork well, then I can see where youâre coming from, though I think I still weakly disagree. For example, if you have an understanding of how these models work under the hood then you are probably better placed to understand which tasks it will be good at and which it might struggle with (this includes really basic stuff, like knowledge cutoffs in training data, and the fact it wonât necessarily remember what you asked it about the day before without a memory feature bolted on, and therefore wonât improve day by day etc). You are also probably in a better position to understand things like which tasks will require higher reasoning, and which wonât (which very often doesnât line up with their perceived human difficulty).
But that aside, I donât think applying generative AI to practical problems is just about prompt engineering or becoming a claude cowork super-user. If you are a business that wants to create an automated workflow where you run a certain prompt over a certain dataset at scale, on the basis of certain triggers, then you are doing something that looks a lot like data engineering (and you are going to need to write code). And if youâre hitting context window limits, youâll have to manage context, by building some kind of agent orchestration/âscaffolding. Thatâs maybe closer to being a brand new skill, but it still requires strong coding ability, combined with the ability to evaluate the performance of a software tool empirically, which is a specific combination of skillsets that data scientists should already have.
Itâs true that with AI coding assistants, people are writing less and less code by hand. But you still need strong coding ability to use AI coding assistants effectively. Software engineering has not yet been fully automated. Thereâs a possible future where that changes soon, and maybe thatâs the future informing your advice in that final section? But my personal feeling would be that in that world, most knowledge jobs probably follow shortly after anyway (if not before), so trying to plan your career development around this possible future seems challenging!
I may be misunderstanding something in the argument, but this potential solution is I think already discussed in detail in Reasons+Persons, where Parfit introduces the repugnant conclusion? He calls it the âlexical valueâ solution, I think? And I donât think this write up addresses any of the strong arguments that Parfit makes against it there?
In Parfitâs original argument, he doesnât just rely on an appeal to intuition that B is better than A. Instead we move to B from A in two steps: first we go to A+ (more people having lives just worth living, but original population unaffected). Parfit claims that A+ is not worse than A. And then we move to world B, where we are now significantly increasing the welfare of already existing people, with only a small drop in the welfare of the original population. Parfit claims that B is better than A+.
If we reject that A+ is not worse than A, then we have to say it is sometimes wrong to create new people, even if their lives are worth living. This seems very strange. If we reject that B is better than A, then we have to believe that a small drop in welfare for people whose welfare is already very high can outweigh a large increase in welfare for people whose welfare is very low. This is the bullet that I think the lexical value solution bites?
But this is a tough bullet to bite! It is the opposite of egalitarian. It says we should prioritise changes in the welfare of already existing high welfare people (whose welfare is above the lexical level) over changes in the welfare of also already existing low welfare people (welfare below the lexical level).
This is all very interesting, thanks for following up!
Something seems especially weird about offsetting your purchase of non-BCC chicken by donating to campaigns to get supermarkets to adopt the BCC.
I think one important consideration missing here: supermarkets respond to campaigners by saying that customers want to buy non-BCC chicken, and they are just doing what their customers want. If you buy non-BCC chicken from them, you make that argument stronger, and the campaignersâ argument weaker.
And I donât think this is necessarily a negligible concern in comparison to the other effects being discussed here, since the mechanism for how your small donation is supposed to help chickens is also by tipping the scales on some corporate campaign and getting a company like a supermarket to make a big change.
I donât deny that my âunlimited time, ink, and paperâ caveat is doing a lot of work in my argument. But we started with a thought experiment that is impossible to implement in practice (simulating a modern digital computer with a pen and paper) so I donât see why my reply canât do the same thing (even if it might require a lot more resources).
I think itâs very unlikely that the human brain requires infinite time and memory to simulate. Even if continuous, you could probably simulate to arbitrary accuracy with a big enough discrete approximation. And the Bekenstein bound suggests there is a finite limit to the amount of information that can exist within a given volume.
As for whether my speed analogy works, I still think it does. Sure, if you pick a frame of reference in which you are stationary, then you continue to have experiences at the normal rate. But that wasnât the frame of reference I was using. I was working in the frame of reference of someone back on Earth, which is an equally valid frame of reference. In those coordinates, every physical process in your brain is getting slowed down (electrical impulses are travelling slower from one side of your brain to the other, chemical reactions are slowing down, etc) and you are having experiences at a slower rate.
If the human brain operates according the known laws of physics, then in principle your brain could be simulated with a pen and paper (at least given unlimited time, ink, and paper), and it would behave identically to the real thing (it would talk and think like you and have all your opinions).
Suppose this was all that existed of you, and your real brain never had existed. Would that mean that you never existed as a conscious being, despite all your thoughts and utterances still being a part of the world? That seems like a much more counter intuitive conclusion to me than biting the bullet on pen+paper simulations having the potential for consciousness.
I donât get why the âmoment of experience taking a thousand yearsâ thing is supposed to be so weird? If we slowed down all the processes in your brain then moments of experience would take longer in physical time. Thatâs not an argument against your consciousness being real. And this isnât a hypothetical. We can literally do that by sending you on a spaceship close to the speed of light, and thatâs exactly what would happen!
That makes a lot of sense, thanks.
Iâm sorry youâve said you regret your engagement, since Iâve found your comments helpful (the link to AISLEâs OpenSSL zero days has shifted my view on this a fair bit).
I guess this whole discussion does just feel like a classic example of âAll debates are bravery debatesâ.
Thanks for the detailed reply, I understand your point clearly now I think!
But $20,000 for *all* of the OpenBSD bugs (not just the published ones) doesnât sound like that much to spend on inference compute to me. If AISLE could have spent the same and made an equally impressive announcement, unearthing enough bugs at once that government ministers around the world start issuing statements about it, then shouldnât they have been able to find the investors to fund that? That would have been incredible publicity for them.
The crux for me seems to be whether they have made equally impressive announcements, as you suggest they might have done. Maybe theyâre just worse at marketing. I donât know enough to evaluate that claim properly, but that does seem the relevant question here: have Anthropic been able to use Mythos to go significantly beyond what the best harnesses could already achieve with existing models for the same inference spend? I thought the answer was a clear yes, and I didnât find the original linked AISLE writeup very convincing at all. Your comment has made me more uncertain, but has still not convinced me, and Iâd be really interested to read something more in depth on that question. (Maybe we also would disagree about what the word âsignificantlyâ means here, since I guess you are acknowledging it probably represents some improvement).
(Also, Iâd push back a bit on your characterization of AI progress. I agree the scaffolding is extremely important, but in my experience the âparadigm shiftsâ in capability over the last two and a half years Iâve been working with them have come from the models)
(And extra comment: the fact that cybersecurity capabilities might not imply imminent superintelligence takeoff seems an entirely independent point that I donât necessarily disagree with)
On the take by AISLE, maybe Iâm missing something here, but if their headline claim was correct (that the harness is more important than the model), shouldnât they have been able to find the vulnerabilities that Anthropic hasnât published? Or find hundreds more similarly impactful ones?
Re-discovering the ones Anthropic had already published seems much less impressive, because there are lots of ways to cheat, and from their write up it sounded to me like they were essentially admitting that they had cheated.
Of course Anthropic could be lying about the existence or significance of the vulnerabilities they havenât published. But they have committed in advance to what those vulnerabilities are (I think they have already made some kind of cryptographic commitment to their unpublished write ups..?) which seems impressive to me.
Either they have used the new model to find significant vulnerabilities in every major OS and browser that are too dangerous to be released, or they havenât. If they have, it seems genuinely scary and impressive (not just marketing hype), because Iâm not aware people working on fancy harnessing have had similar results (or have they?) And if they havenât, then itâs a very weird marketing ploy, because theyâre going to get found out very quickly!
I think this misunderstands what people mean when they compare arguments about the importance of AI safety to a Pascalâs wager.
Pascalâs wager refers to situations where a tiny probability of enormous value seemingly leads to ridiculous conclusions if you try to do naive expected value calculations with it. When people say that strong longtermism is a Pascalâs wager, the âsmall probabilityâ they are talking about is not the probability of extinction, which as you point out, is significant. The small probability is the probability that the future will contain âseptillions of future sapientsâ. That is the probability that is small. And it gets even smaller if the probability of extinction soon is high! So a large probability of extinction this century makes the Pascalâs wager comparison more relevant as a critique of strong longtermism, not less. It is multiplying this small probability by the value of those septillions of potential âsapientsâ that gives you the astronomical value that says existential risk reduction should almost automatically dominate our concerns.
I think youâre completely right to point out that people should care a lot about things which might carry a 10% chance of causing human extinction, even ignoring their stance on longtermism. But some people believe that existential risk has astronomically more value than just the impact it will have on the next few generations, and that therefore tiny changes in the probability of existential risk almost automatically trump any other concern, however small those changes are. When people talk about Pascalâs wager in the context of strong longtermism or AI safety, I think it is this claim that they are challenging, not the claim that we should care about extinction at all. And that criticism is just as valid, actually more valid, if the probability of extinction from AI safety is high (though I of course agree that if there are people who use the Pascalâs Wager argument to dismiss all work on AI risk then they are making a serious mistake).
I agree with your title, but I donât think negative utilitarianism is the answer. I like Toby Ordâs essay on this, âWhy Iâm Not a Negative Utilitarianâ: https://ââwww.amirrorclear.net/ââacademic/ââideas/âânegative-utilitarianism/ââ
On your argument about tradeoffs, people make choices all the time where they accept some very small risk of some very severe suffering in order to increase their happiness by a modest amount. For example: cycling along a busy road to visit their friend. If you say that no amount of happiness can make up for the trauma of being involved in a serious accident, then it seems like you are forced to say that this choice is wrong. That seems like a strange conclusion to me.
I think itâs difficult to give a complete answer to your question without essentially drafting a fresh entry to the cluelessness critiques competition. I accept that it is not a solution to the general problem to simply say âjust ignore the effects/âcausal pathways youâre clueless aboutâ, because as you point out, in general, we can change the way we group these effects, and that may change what gets ignored. Iirc there is a section arguing this point clearly in digiovannis sequence.
My example was not meant to be taken as saying âlook, there is this obviously correct method for approaching these problems that bracketing disagrees withâ. It was meant to be taken only as âin this very contrived example, bracketingâs recommendations seem to clash with the intuition that motivates bracketing in the first placeâ (maybe I am going wrong by thinking it motivated by intuition)
In my contrived example, there are no other good things and other bad things. The effects are, by assumption, just (splitting positive and negative):
I get a tasty snack right away.
Broken glass left somewhere and injures someone.
Dystopian lock in triggered that otherwise would not have happened.
Dystopian lock in averted that otherwise would have happened.
Effect 1 occurs with probability 1. Effect 2 occurs with probability near 1 (if imprecise, the interval is small and near 1).
Effects 3 and 4 both occur with only very small probability (and are clearly mutually exclusive as well). And the probabilities are imprecise. But the upper limit of each imprecise interval is large enough to produce value in expectation that outweighs 1 and 2. And the intervals overlap, which makes the sign of the total expected value indeterminate.
This is I think a complete specification of the problem (we donât need to worry about the causal pathways?). And my intuition still strongly points to one of these conclusions being acceptable:
It is indeterminate whether we should want this act to happen.
We should prefer that this act does not happen.
But bracketing advocates neither of these, so feels wrong.
If you push me on what it is that lets me group 3 and 4 together so that I can ignore them, rather than grouping 2, 3, and 4 together (as bracketing does) I guess I would point to the degree of imprecision in the probability estimates, which feels like the more relevant factor than which person is feeling the effect.