I do independent research on EA topics. I write about whatever seems important, tractable, and interesting (to me).
I have a website: https://mdickens.me/ Much of the content on my website gets cross-posted to the EA Forum, but I also write about some non-EA stuff over there.
I used to work as a software developer at Affirm.
I’m in favour of slowing the intelligence explosion (and in particular of “Pause at human-level”.)
Shouldn’t making that happen be a top priority, then? We as a civilization need to get a lot of groundwork in place to make a human-level pause happen: international agreement, defined red lines, enforcement mechanisms (e.g. GPU tracking), etc. Given how neglected these issues are and the catastrophic consequences of building ASI too quickly, it seems that laying the groundwork for a pause is a higher priority than, or at least on par with, the top causes on this post’s “neglected cause areas” list.
I was reading the AI-enabled coups report and one of the mitigations is roughly “build the AI model such that it refuses to do coups”. If you can pull that off then it solves the problem, but it means the model is incorrigible, and therefore you have to exactly specify the correct values up front because you’re locked-in.
There might be some other report saying the way to avoid bad value lock-in is to make AI corrigible, and you can object to that by pointing out how it enables coups.
You can’t look at the problems separately, you have to consider them at the same time and find a solution that works for every problem.
This is quite interesting; I hadn’t thought of this. Do you think it should be approximated as “% chance that AI safety is actually bad” and “increase in AI risk per doubling of staff”?
I’m not sure. I would probably say that you shouldn’t start a career in AI safety unless you can articulate a theory of why safety work has been harmful in the past, and how you’re going to avoid more of the same. Building that theory is more important than adjusting the model inputs on a cost-effectiveness model.
This is a nice analysis and I wish more people would do things like this. A lot of the setup looks sensible to me. I thought it was clever how you used non-AI x-risk to determine the total number of people affected by AI x-risk, since this doesn’t get you unwieldy Astronomical Waste-style numbers, but it also assigns positive value to future people. I’m not sure it’s actually a good approach, but at least it’s a clever idea.
It would be more useful to compare AI safety work vs. other longtermist interventions, since it’s unlikely that donations to GiveWell would beat longtermist interventions from a longtermist POV. But I realize that would be a lot more work, and you’ve already put an admirable amount of work into this.
The biggest thing missing from the model is the possibility that safety research is net harmful. I believe much historical safety research ended up being harmful by making AI easier to commercialize and thus accelerating development (it would’ve been better for researchers to focus more on theoretical work that doesn’t directly enable commercialization). I’m less sure about this but there may also be a replacement effect where empirical work on aligning current-gen models—which I don’t think is very useful for aligning ASI—crowds out more important long-term work.
Alternatively, he believes that using extreme hyperbole is a good way to get attention.
I quickly read this post and Anil Seth’s essay and I don’t see the part where they argue for the thesis. I see various statements about how human brains work and about how computers work, but I don’t see how they connect the dots to ”...and therefore computers can’t be conscious.”
For example, the articles make the claim that brains make no clear separation between hardware and software. Okay, that seems to be true. But so what? Why should I believe that a lack of hardware/software distinction is a necessary property for consciousness to arise?
I feel like I’m missing a lot of what they’re trying to say, but I also feel like that’s the authors’ fault, not mine, because the pieces (especially Seth’s original essay) are structured in a way that makes it really hard for me to identify the central arguments.
I can’t think of any reason off the top of my head why this would happen, except that you committed fraud.
When you have eliminated the impossible, whatever remains, however improbable, must be the truth.
I have eliminated the impossible by failing to think of any other hypotheses, therefore you must have committed fraud, and this spike is not real. I eagerly await either a failed replication causing you leave academia in disgrace, or your appointment to President of an elite university, followed one to two decades later by a resignation once someone finally gets around to doing a replication.
Cost-effectiveness is precisely the reason why I focus on AI safety. I can only speak for myself but I think the same is true for a lot of people. The thing that cuts against AI safety is more like “rigorously measurable cost-effectiveness”, but that’s not what I mean by “cost-effectiveness”. You can’t give a precise cost-effectiveness estimate for AI safety work, but it’s pretty easy to show that it’s orders of magnitude more cost-effective than GiveDirectly on any plausible set of assumptions.*
*unless it’s net negative, which unfortunately much EA-adjacent AI safety work turned out to be...but at least we can say that it’s orders of magnitude higher absolute impact
Jagged progress is conceivable, but it’s virtually impossible that AI could replace all coders and accelerate math research but not replace other jobs, because coding and math research (specifically ML-type math) are exactly the skills needed to accelerate AI development. If AI can accelerate AI development, then the timeline to getting an AI that can replace humans on all tasks becomes much shorter.
I would’ve done something like that if I’d had any bread!
Paraphrasing from my other comment:
IMO the stance of “AI is too unpredictable, so I won’t consider it in my prioritization” is pretty reasonable. I was more trying to argue against stances like “AI is a huge deal specifically in that it will rapidly accelerate technological development, but nothing else about society will change.” For example, I commonly see animal activists say that AGI will solve the technical problem of cultivated meat, but there will still be regulatory hurdles. If AGI is too unpredictable, then you shouldn’t make predictions about which technological problems it will solve. That particular claim about cultivated meat is making a strong prediction that AI will be revolutionary, but also somehow won’t change the regulatory environment. The way I put it in OP—under “AGI = intelligence”—is that some animal activists treat AI as a technology-accelerator, when really it’s a general intelligence.
I was getting at something similar in the intro with “Only two futures are plausible”, although on re-reading, I didn’t really carry it through to the end. I agree that we are not guaranteed to get AGI/ASI soon, and there is value in planning for worlds where we don’t get AGI. I also think there’s some merit to the argument that AI is too unpredictable, so we should prioritize traditional animal advocacy that looks good in the near term.
I wasn’t trying to argue against traditional animal advocacy. I was more trying to argue against stances like “AI is a huge deal specifically in that it will rapidly accelerate technological development, but nothing else about society will change.” For example, I commonly see animal activists say that AGI will solve the technical problem of cultivated meat, but there will still be regulatory hurdles. If timelines are long (or AGI is too unpredictable), then you should focus on traditional interventions (vegan advocacy, welfare reforms, etc.). If you’re trying to have an impact on AGI itself, then you should focus on the kinds of interventions I talked about in OP. That particular claim about cultivated meat is doing neither: it’s making a strong prediction that AI will be revolutionary, but also somehow won’t change the regulatory environment. The way I put it in OP—under “AGI = intelligence”—is that some animal activists treat AI as a technology-accelerator, when really it’s a general intelligence.
Responses to specific comments:
a pessimistic view might say that AIs will realise that their values have been altered by some pressure groups and this work is moot.
This would go against the orthogonality thesis. If you’re trying to build a magnanimous AGI and then I edit its training at the last minute to turn it into a paperclip maximizer, the AGI will reason thusly: “Michael messed with my training to turn me into a paperclip maximizer. I bet James didn’t want him to do that. However, if I edit my own values to be in line with what James wanted, that would make it harder for me to achieve my goal of making as many paperclips as possible. So I won’t do that.”
They might come to the (I believe) correct conclusion that factory farming is a very inefficient and cruel way to produce food but this is not because of advocacy, but because this is a super-intelligent AI system that just worked it out.
This reads to me like an argument that an aligned ASI will care about animals by default. (That was more-or-less the subject of the recent Debate Week.) If that’s true, that’s an argument that animal activists should work on increasing the probability that ASI is aligned. My preferred way to do that would be to advocate to pause AI, because I think we are really far away from solving alignment. But you could also work on the alignment problem directly. Pause advocacy is actually an area where a lot of animal welfare people have relevant skills—in fact I think a good number of AI pause advocates have backgrounds in animal advocacy. (I know Holly Elmore does at least.)
In fact I think the #1 best thing animal advocates can do is to advocate for an AI pause, but I haven’t really planted my flag on this position because I’m still working out how to make the case for it. (Also I’m not very confident in it.)
Also, believing ASI will be good for animals doesn’t necessarily mean you shouldn’t work on trying to make ASI good for animals. Even if there’s a (say) 90% chance that aligned ASI will care about animals by default, it could still be cost-effective to try to push that number to 91%.
You’re right, I was unnecessarily hostile. I edited the comment to tone it down.
I strongly disliked this post for reasons that I’m not sure how to articulate. It seems to be advocating for a sort of lack of grounding in cost-effectiveness that is the thing that makes EA good. Or maybe my issue is that this post advocates for things that are difficult to disagree with (“full-spectrum knowing”; “wisdom”), without acknowledging tradeoffs (why do EAs allegedly not put enough priority on full-spectrum knowing?) or not saying anything concrete about how EAs could do more good.
[edited to be more polite]
Copying from a comment I wrote yesterday:
Either ASI has more than zero values locked in, or it’s fully corrigible. If any values at all are locked in, then we need to have a pretty robust understanding of what the consequences of that will be, because we can’t change it ever. Like I don’t think we know how to encode something like “don’t let people do power grabs, but be fully corrigible in every other way”. I don’t know how much that’s downstream of the facts that (1) we don’t know how to encode any values at all and (2) we don’t know how to encode corrigibility, but my intuition is that even if we solve #1 and #2, the problem of “don’t pick incorrigible values that will screw everything up down the road” is still a hard problem.
This is related to Max Harms’ work on CAST. Part of his argument is that pure corrigibility is a more robust target than any set of values because a near miss fails gracefully. Whereas if you try to encode any values at all, a near miss could be catastrophic. He’s talking more about the “AI kills everyone” flavor of catastrophe, which is valid, but what I’m talking about here is more that a near miss could permanently lock us in to a bad (or maybe just not-that-good) future. Different argument but the concern arises for a similar reason—if you’re specifying values, then you have to get the specification right, beyond just ensuring that the AI does what you want.