For 5, the survey (https://arxiv.org/pdf/1705.08807.pdf) sort of ends all discussion about AI timelines. Not that it’s necessarily right, just that no one is in a position to second-guess it.
For 7, we can learn quite a bit from working on long term causes, and WASR is an example of that: it’s stuff that won’t be implemented any time soon, but we can gain feedback from the baby steps. The same thing has applied to some AI work.
Also, it seems to me that the kind of expertise here is highly domain-specific, and the lessons learned in one domain probably won’t help elsewhere. I suppose that short term causes let you perform more trials after observing initial results, at least.
For 8, nontechnical people can work on political issues with long-term implications.
Lists of 10 are always fishy because the author is usually either stretching them out with poor reasons to make it to 10, or leaving out good reasons to keep it at 10. Try not to get attached to the number :)
I do agree WASR seems pretty tractable and the near-term learning value is pretty high even if we don’t have a good idea of the long-term feasibility yet. I think it’s promising, but I could also see it being ruled out as impactful, and I feel like we could have a good answer in a few years.
I don’t have a good sense yet on whether something like AI research has a similar feel. If it did, I’d feel more excited about it.
For 5, the survey (https://arxiv.org/pdf/1705.08807.pdf) sort of ends all discussion about AI timelines. Not that it’s necessarily right, just that no one is in a position to second-guess it.
I don’t follow what you mean by “ends all discussion.”
Even if AI development researchers had a consensus opinion about AI timelines (which they don’t), one could still disagree with the consensus opinion.
I suspect AI dev researcher timeline estimates vary a lot depending on whether the survey is conducted during an AI boom or AI winter.
Well, you might disagree, but you’d have to consider yourself likely to be a better predictor than most AI experts.
The lack of consensus doesn’t really change the point because we are looking at a probability distribution either way.
Booms/winters are well known among researchers, they are aware of how it affects the field so I think it’s not so easy to figure out if they’re being biased or not.
I think it’s important to hold “AI development research” and “AI timeline prediction-making” as two separate skillsets. Expertise in one doesn’t necessarily imply expertise in the other (though there’s probably some overlap).
Any good model of the quality of AI dev researcher timeline opinions needs to be able to explain why AI safety was considered a joke by the field for years, and only started to be taken seriously by (some) AI dev researchers after committed advocacy from outsiders.
I think it’s important to hold “AI development research” and “AI timeline prediction-making” as two separate skillsets. Expertise in one doesn’t necessarily imply expertise in the other (though there’s probably some overlap).
OK, that’s true. The problem is, it’s hard to tell if you are better at predicting timelines.
Any good model of the quality of AI dev researcher timeline opinions needs to be able to explain why AI safety was considered a joke by the field for years, and only started to be taken seriously by (some) AI dev researchers after committed advocacy from outsiders.
I think that’s a third issue, not a matter of timeline opinions either.
I think that’s a third issue, not a matter of timeline opinions either.
Seems relevant in that if you surveyed timeline opinions of AI dev researchers 20 years ago, you’d probably get responses ranging from “200 years out” to “AGI? That’s apocalyptic hogwash. Now, if you’d excuse me...”
I don’t know which premise here is more greatly at odds with the real beliefs of AI researchers—that they didn’t worry about AI safety because they didn’t think that AGI would be built, or that there has ever been a time when they thought it would take >200 years to do it.
For 5, the survey (https://arxiv.org/pdf/1705.08807.pdf) sort of ends all discussion about AI timelines. Not that it’s necessarily right, just that no one is in a position to second-guess it.
For another relevant reason to think less about the future, take a look at this. https://web.stanford.edu/~chadj/IdeaPF.pdf
For 7, we can learn quite a bit from working on long term causes, and WASR is an example of that: it’s stuff that won’t be implemented any time soon, but we can gain feedback from the baby steps. The same thing has applied to some AI work.
Also, it seems to me that the kind of expertise here is highly domain-specific, and the lessons learned in one domain probably won’t help elsewhere. I suppose that short term causes let you perform more trials after observing initial results, at least.
For 8, nontechnical people can work on political issues with long-term implications.
Lists of 10 are always fishy because the author is usually either stretching them out with poor reasons to make it to 10, or leaving out good reasons to keep it at 10. Try not to get attached to the number :)
I do agree WASR seems pretty tractable and the near-term learning value is pretty high even if we don’t have a good idea of the long-term feasibility yet. I think it’s promising, but I could also see it being ruled out as impactful, and I feel like we could have a good answer in a few years.
I don’t have a good sense yet on whether something like AI research has a similar feel. If it did, I’d feel more excited about it.
I don’t follow what you mean by “ends all discussion.”
Even if AI development researchers had a consensus opinion about AI timelines (which they don’t), one could still disagree with the consensus opinion.
I suspect AI dev researcher timeline estimates vary a lot depending on whether the survey is conducted during an AI boom or AI winter.
Well, you might disagree, but you’d have to consider yourself likely to be a better predictor than most AI experts.
The lack of consensus doesn’t really change the point because we are looking at a probability distribution either way.
Booms/winters are well known among researchers, they are aware of how it affects the field so I think it’s not so easy to figure out if they’re being biased or not.
I think it’s important to hold “AI development research” and “AI timeline prediction-making” as two separate skillsets. Expertise in one doesn’t necessarily imply expertise in the other (though there’s probably some overlap).
Any good model of the quality of AI dev researcher timeline opinions needs to be able to explain why AI safety was considered a joke by the field for years, and only started to be taken seriously by (some) AI dev researchers after committed advocacy from outsiders.
OK, that’s true. The problem is, it’s hard to tell if you are better at predicting timelines.
I think that’s a third issue, not a matter of timeline opinions either.
Seems relevant in that if you surveyed timeline opinions of AI dev researchers 20 years ago, you’d probably get responses ranging from “200 years out” to “AGI? That’s apocalyptic hogwash. Now, if you’d excuse me...”
I don’t know which premise here is more greatly at odds with the real beliefs of AI researchers—that they didn’t worry about AI safety because they didn’t think that AGI would be built, or that there has ever been a time when they thought it would take >200 years to do it.