= 27d567bc2d48d9f40e9ccc20ea370b15)
jsteinhardt comments on The $125,000 Summer Singularity Challenge - Less Wrong
You are viewing a comment permalink. View the original post to see all comments and the full post content.
= 783df68a0f980790206b9ea87794c5b6)
Loading…= b715d02e85f7b3b4ae65ca3d3e450868)
Subscribe to RSS Feed
Powered by Reddit
= f037147d6e6c911a85753b9abdedda8d)
You are viewing a comment permalink. View the original post to see all comments and the full post content.
Comments (259)
Even if this is so, there is tons of evidence that humans suck at reasoning about such large numbers. If you want to make an extraordinary claim like the one you made above, then you need to put forth a large amount of evidence to support it. And on such a far-mode topic, the likelihood of your argument being correct decreases exponentially with the number of steps in the inferential chain.
I only skimmed through the video, but assuming that the estimates at 11:36 are what you're referring to, those numbers are both seemingly quite high and entirely unjustified in the presentation. It also overlooks things like the fact that utility doesn't scale linearly in number of lives saved when calculating the benefit per dollar.
Whether or not those numbers are correct, presenting them in their current form seems unlikely to be very productive. Likely either the person you are talking to already agrees, or the 8 lives figure triggers an absurdity heuristic that will demand large amounts of evidence. Heck, I'm already pretty familiar with the arguments, and I still get a small amount of negative affect whenever someone tries to make the "donating to X-risk has <insert very large number> expected utility".
I don't think anyone on LW disagrees that reducing xrisk substantially carries an extremely high utility. The points of disagreement are over whether SIAI can non-trivially reduce xrisk, and whether they are the most effective way to do so. At least on this website, this seems like the more productive path of discussion.
Keep in mind that estimation is the best we have. You can't appeal to Nature for not having been given a warning that meets a sufficient standard of rigor. Avoiding all actions of uncertain character dealing with huge consequences is certainly a bad strategy. Any one of such actions might have a big chance of not working out, but not taking any of them is guaranteed to be unhelpful.
From a Bayesian point of view, your prior should place low probability on a figure like "8 lives per dollar". Therefore, lots of evidence is required to overcome that prior.
From a decision-theoretic point of view, the general strategy of believing sketchy (with no offense intended to Anna; I look forward to reading the paper when it is written) arguments that reach extreme conclusions at the end is a bad strategy. There would have to be a reason why this argument was somehow different from all other arguments of this form.
If there were tons of actions lying around with similarly huge potential positive consequences, then I would be first in line to take them (for exactly the reason you gave). As it stands, it seems like in reality I get a one-time chance to reduce p(bad singularity) by some small amount. More explicitly, it seems like SIAI's research program reduces xrisk by some small amount, and a handful of other programs would also reduce xrisk by some small amount. There is no combined set of programs that cumulatively reduces xrisk by some large amount (say > 3% to be explicit).
I have to admit that I'm a little bit confused about how to reason here. The issue is that any action I can personally take will only decrease xrisk by some small amount anyways. But to me the situation feels different if society can collectively decrease xrisk by some large amount, versus if even collectively we can only decrease it by some small amount. My current estimate is that we are in the latter case, not the former --- even if xrisk research had unlimited funding, we could only decrease total xrisk by something like 1%. My intuitions here are further complicated by the fact that I also think humans are very bad at estimating small probabilities --- so the 1% figure could very easily be a gross overestimate, whereas I think a 5% figure is starting to get into the range where humans are a bit better at estimating, and is less likely to be such a bad overestimate.
My prior contains no such provisions; there are many possible worlds where tiny applications of resources have apparently disproportionate effect, and from the outside they don't look so unlikely to me.
There are good reasons to be suspicious of claims of unusual effectiveness, but I recommend making that reasoning explicit and seeing what it says about this situation and how strongly.
There are also good reasons to be suspicious of arguments involving tiny probabilities, but keep in mind: first, you probably aren't 97% confident that we have so little control over the future (I've thought about it a lot and am much more optimistic), and second, that even in a pessimistic scenario it is clearly worth thinking seriously about how to handle this sort of uncertainty, because there is quite a lot to gain.
Of course this isn't an argument that you should support the SIAI in particular (though it may be worth doing some information-gathering to understand what they are currently doing), but that you should continue to optimize in good faith.
Can you clarify what you mean by this?
Only that you consider the arguments you have advanced in good faith, as a difficulty and a piece of evidence rather than potential excuses.
I'm glad you agree.
I'd be very appreciative to hear if you know of someone doing more.
Over the coming months I'm going to be doing an investigation of the non-profits affiliated with the Nuclear Threat Initiative with a view toward finding x-risk reduction charities other than SIAI & FHI. I'll report back what I learn but it may be a while.
I'm under the impression that nuclear war doesn't pose an existential risk. Do you disagree? If so, I probably ought to make a discussion post on the subject so we don't take this one too far off topic.
My impression is that the risk of immediate extinction due to nuclear war is very small but that a nuclear war could cripple civilization to the point of not being able to recover enough to affect a positive singularity; also it would plausibly increase other x-risks - intuitively, nuclear war would destabilize society, and people are less likely to take safety precautions in an unstable society when developing advanced technologies than they otherwise would be. I'd give a subjective estimate of 0.1% - 1% of nuclear war preventing a positive singularity.
Do you mean:
Good question. My intended meaning was the second of the meanings that you listed "the probability of a positive singularity is 0.1%-1% lower than the probability of a positive singularity given no nuclear war." Would be interested to hear any thoughts that you have about these things.
I can't think of a mechanism through which recovery would become long-term impossible, but maybe there is one. People taking fewer safety precautions in a destabilized society does sound plausible. There are probably a number of other, similarly important effects of nuclear war on existential risk to take into account. Different technologies (IA, uploading, AGI, Friendliness philosophy) have different motivations behind them that would probably be differently affected by a nuclear war. Memes would have more time to come closer to some sort of equilibrium in various relevant groups. To the extent that there are nontrivial existential risks not depending on future technology, they would have more time to strike. Catastrophes would be more psychologically salient, or maybe the idea of future nuclear war would overshadow other kinds of catastrophe. Power would be more in the hands of those who weren't involved in the nuclear war.
In any case, the effect of nuclear war on existential risk seems like a nontrivial question that we'd have to have a better idea about before we could decide that resources are better spent on nuclear war prevention than something else. To make things more complicated, it's possible that preventing nuclear war would on average decrease existential risk but that a specific measure to prevent nuclear war would increase existential risk (or vice versa), because the specific kinds of nuclear war that the measure prevents are atypical.
The number and strength of reasons we see one way or the other may depend more on time people have spent searching specifically for reasons for/against than on what reasons exist. The main reason to expect an imbalance there is that nuclear war causes huge amounts of death and suffering, and so people will be motivated to rationalize that it will also be a bad thing according to this mostly independent criterion of existential risk minimization; or people may overcorrect for that effect or have other biases for thinking nuclear war would prevent existential risk. To the extent that our misgivings about failing to do enough to stop nuclear war have to do with worries that existential risk reduction may not outweigh huge present death and suffering, we'd do better to acknowledge those worries than to rationalize ourselves into thinking there's never a conflict.
Without knowing anything about specific risk mitigation proposals, I would guess that there's even more expected return from looking into weird, hard-to-think-about technologies like MNT than from looking into nuclear war, because less of the low-hanging fruit there would already have been picked. But more specific information could easily overrule that presumption, and some people within SingInst seem to have pretty high estimates of the return from efforts to prevent nuclear war, so who knows.
Thanks for your thoughtful comment.
I have little idea of how likely it is but a nuclear winter could seriously hamper human mobility.
Widespread radiation would further hamper human mobility.
Redeveloping preexisting infrastructure could require natural resources on of order of magnitude comparable to the infrastructure that we have today. Right now we have the efficient market hypothesis to help out with natural resource shortage, but upsetting the trajectory of our development could exacerbate the problem.
Note that a probability of 0.1% isn't so large (even taking account all of the other things that could interfere with a positive singularity).
Reasoning productively about the expected value of these things presently seems to me to be too difficult (but I'm open to changing my mind if you have ideas).
With the exception of natural resource shortage (which I mentioned above) I doubt that this is within an order of magnitude of significance of other relevant factors provided that we're talking about a delay on the order of fewer than 100 years (maybe similarly for a delay of 1000 years; I would have to think about it).
Similarly, I doubt that this would be game-changing.
These seem worthy of further contemplation - is the development of future technologies more likely to go in Australia than in the current major powers, etc.
This seems reasonable. As I mentioned, I presently attach high expected x-risk reduction to nuclear war prevention but my confidence is sufficiently unstable at present so that the value devoting resources to gather more information outweighs the value of donating to nuclear war reduction charities.
Yes. In the course of researching nuclear threat reduction charities I hope to learn what options are on the table.
On the other hand there may not be low hanging fruit attached to thinking about weird, hard-to-think-about technologies like MNT. I do however plan on looking into the Foresight Institute.
Thanks for clarifying and I hope your research goes well. If I'm not mistaken, you can see the 0.1% calculation as the product of three things: the probability nuclear war happens, the probability that if it happens it's such that it prevents any future positive singularities that otherwise would have happened, and the probability a positive singularity would otherwise have happened. If the first and third probabilities are, say, 1/5 and 1/4, then the answer will be 1/20 of the middle probability, so your 0.1%-1% answer corresponds to a 2%-20% chance that if a nuclear war happens then it's such that it prevents any future positive singularities that would otherwise have happened. Certainly the lower end and maybe the upper end of that range seem like they could plausibly end up being close to our best estimate. But note that you have to look at the net effect after taking into account effects in both directions; I would still put substantial probability on this estimate ending up effectively negative, also. (Probabilities can't really go negative, so the interpretation I gave above doesn't really work, but I hope you can see what I mean.)
Thanks for the clarification on the estimate. Unhappy as it makes me to say it, I suspect that nuclear war or other non-existential catastrophe would overall reduce existential risk, because we'd have more time to think about existential risk mitigation while we rebuild society. However I suspect that trying to bring nuclear war about as a result of this reasoning is not a winning strategy.
Building society the first time around, we were able to take advantage of various useful natural resources such as relatively plentiful coal and (later) oil. After a nuclear war or some other civilization-wrecking catastrophe, it might be Very Difficult Indeed to rebuild without those resources at our disposal. It's difficult enough even now, with everything basically still working nicely, to see how to wean ourselves off fossil fuels, as for various reasons many people think we should do. Now imagine trying to build a nuclear power industry or highly efficient solar cells with our existing energy infrastructure in ruins.
So it looks to me as if (1) our best prospects for long-term x-risk avoidance all involve advanced technology (space travel, AI, nanothingies, ...) and (2) a major not-immediately-existential catastrophe could seriously jeapordize our prospects of ever developing such technology, so (3) such a catastrophe should be regarded as a big increase in x-risk.
I've heard arguments for and against "it might turn out to be too hard the second time around". I think overall that it's more likely than not that we would eventually succeed in rebuilding a technological society, but that's the strongest I could put it, ie it's very plausible that we would never do so.
If enough of our existing thinking survives, the thinking time that rebuilding civilization would give us might move things a little in our favour WRT AI++, MNT etc. I don't know which side does better on this tradeoff. However I seriously doubt that trying to bring about the collapse of civilization is the most efficient way to mitigate existential risk.
Also, and I hate to be this selfish about it but there it is, if civilization ends I definitely die either way, and I'd kind of prefer not to.
We have a huge mountain of coal, and will do for the next hundred years or so. Doing without doesn't seem very likely.
How easily accessible is that coal to people whose civilization has collapsed, taking most of the industrial machinery with it? (That's a genuine question. Naively, it seems like the easiest-to-get-at bits would have been mined out first, leaving the harder bits. How much harder they are, and how big a problem that would be, I have no idea.)
It's probably fair to say that some of the low hanging fossil fuel fruit have been taken.
A more likely result: the religious crazies will take over, and they either don't think existential risk can exist (because God would prevent them) or they think preventing existential risk would be blasphemy (because God ought be allowed to destroy us). Or they even actively work to make it happen and bring about God's judgmenent.
And then humanity dies, because both denying and embracing existential risk causes it to come nearer.
Technical challenges? Difficulty in coordinating? Are there other candidate setbacks?
It may be highly unproductive to think about advanced future technologies in very much detail before there's a credible research program on the table on account of the search tree involving dozens of orders of magnitude. I presently believe in this to be the case.
I do think that we can get better at some relevant things at present (learning how to obtain as accurate as realistically possible predictions about probable government behaviors, etc.) and that all else being equal we could benefit from more time thinking about these things rather than less time.
However, it's not clear to me that the time so gained would outweigh a presumed loss in clear thinking post-nuclear war and I currently believe that the loss would be substantially greater than the gain.
As steven0461 mentioned, "some people within SingInst seem to have pretty high estimates of the return from efforts to prevent nuclear war." I haven't had a chance to talk about this with them in detail; but it updates me in the direction of attaching high expected value reduction to nuclear war risk reduction.
My positions on these points are very much subject to change with incoming information.
How much detail is too much?
Well for instance, certain approaches to AGI are more likely to lead to something friendly than other approaches are. If you believe that approach A is 1% less likely to lead to a bad outcome than approach B, then funding research in approach A is already compelling.
In my mind, a well-reasoned statistical approach with good software engineering methodologies is the mainstream approach that is least likely to lead to a bad outcome. It has the advantage that there is already a large amount of related research being done, hence there is actually a reasonable chance that such an AGI would be the first to be implemented. My personal estimate is that such an approach carries about 10% less risk than an alternative approach where the statistics and software are both hacked together.
In contrast, I estimate that SIAI's FAI approach would carry about 90% less risk if implemented than a hacked-together AGI. However, I assign very low probability to SIAI's current approach succeeding in time. I therefore consider the above-mentioned approach more effective.
Another alternative to SIAI that doesn't require estimates about any specific research program would be to fund the creation of high-status AI researchers who care about Friendliness. Then they are free to steer the field as a whole towards whatever direction is determined to carry the least risk, after we have the chance to do further research to determine that direction.
I don't understand what you mean by "10% less risk". Do you think any given project using "a well-reasoned statistical approach with good software engineering methodologies" has at least 10% chance of leading to a positive Singularity? Or each such project has a P*0.9 probability of causing an existential disaster, where P is the probability of disaster of a "hacked together" project. Or something else?
Sorry for the ambiguity. I meant P*0.9.
You said "I therefore consider the above-mentioned approach more effective.", but if all you're claiming is that the above mentioned approach ("a well-reasoned statistical approach with good software engineering methodologies") has a P*0.9 probability of causing an existential disaster, and not claiming that it has a significant chance of causing a positive Singularity, then why do you think funding such projects is effective for reducing existential risk? Is the idea that each such project would displace a "hacked together" project that would otherwise be started?
EDIT: I originally misinterpreted your post slightly, and corrected my reply accordingly.
Not quite. The hope is that such a project will succeed before any other hacked-together project succeeds. More broadly, the hope is that partial successes using principled methodologies will convince them to be more widely adopted in the AI community as a whole, and more to the point that a contingent of highly successful AI researchers advocating Friendliness can change the overall mindset of the field.
The default is a hacked-together AI project. SIAI's FAI research is trying to displace this, but I don't think they will succeed (my information on this is purely outside-view, however).
An explicit instantiation of some of my calculations:
SIAI approach: 0.1% chance of replacing P with 0.1P Approach that integrates with the rest of the AI community: 30% chance of replacing P with 0.9P
In the first case, P is basically staying constant, in the second case it is being replaced with 0.97P.
I noticed you didn't name anybody. Did you have specific programs or people in mind?
We already seem to (roughly) agree on probabilities.
The only specific plan I have right now is to put myself in a position to hire smart people to work on this problem. I think the most robust way to do this is to get a faculty position somewhere, but I need to consider the higher relative efficiency of corporations over universities some more to figure out if it's worthwhile to go with the higher-volatility route of industry.
Also, as Paul notes, I need to consider other approaches to x-risk reduction as well to see if I can do better than my current plan. The main argument in favor of my current plan is that there is a clear path to the goal, with only modest technical hurdles and no major social hurdles. I don't particularly like plans that start to get fuzzier than that, but I am willing to be convinced that this is irrational.
EDIT: To be more explicit, my current goal is to become one of said high-status AI researchers. I am worried that this is slightly self-serving, although I think I have good reason to believe that I have a comparative advantage at this task.
That seems more of an alternative within SIAI than an alternative to SIAI. With more funding, their Associate Research Program can promote the importance of Friendliness and increase the status of researchers who care about it.
Woah, woah! What! Since when?
Unless you mean "scope insensitivity"?
Well, sure, the absurdity heuristic is terrible.
Why would it scale linearly? I agree that is scales linearly over relatively small regimes (on the order of millions of lives) by fungibility, but I see no reason why that needs to be true for trillions of lives or more (and at least some reasons why it can't scale linearly forever).
Re-read the context of what I wrote. Whether or not the absurdity heuristic is a good heuristic, it is one that is fairly common among humans, so if your goal is to have a productive conversation with someone who doesn't already agree with you, you shouldn't throw out such an ambitious figure without a solid argument. You can almost certainly make whatever point you want to make with more conservative numbers.