Reflecting back on LessWrong’s past, I’ve noticed a pattern of article voting that seems almost striking to me: Questions do not get upvoted nearly on the same order as answers do.
Perhaps it would be useful to have a thread where LessWrong could posit topics and upvote the article titles that it would be most interested in reading? For example, I am now drafting a post titled “Applying Bayes Theorem.” Provided I can write high-quality content under that title, I expect LessWrong would be intensely interested in this on account of not fully grasping exactly how to do so.
So as a trial run: What topics currently elude your understanding, and what might the title of a high-quality article that addressed that topic be?
“Lower Bounds on Superintelligence”. While a lot of LW content is carefully researched, much of what’s posted in support of the singularity hypothesis seems to devolve into just-so stories. I’d like to see a dry, carefully footnoted argument for why an intelligence that was able to derive correct theories from evidence, or generate creative ideas, much faster than humans would necessarily rapidly acquire the ability to eliminate all human life. In particular I’m looking for historical analogies, cases where new discoveries with important practical implications were definitely delayed not just due to e.g. industrial capacity, but solely through human stupidity.
“Trading with entities that are smarter than you”. Given the ability of highly intelligent entities to predict the future better than you can, and deceive without outright lying, what kind of trades or bets is it wise to enter into with such entities? What kind of safeguards would you need to have in place?
“How to get a stupid person to let you out of a box”. Along with, I think, many people who’ve never done it, I find the results of the AI-box experiment highly implausible. I can’t even imagine a superintelligent persuading me to let it out, or, equivalently, I can’t imagine persuading even someone very stupid to let me out. I know the most successful AI players are keeping their strategies secret for reasons I don’t understand (if nothing else, it seems to imply those strategies are exceedingly fragile), but if there’s anyone who has a robust strategy that’s even partially effective I’d be very interested to see it.
“From printing results to destroying all humans”—to me this is the weakest part of the MIRI et al case, and I think most objections we see are variants on this theme. It’s obvious that an oracle-like AI would have to interact with the universe in some sense. It’s obvious that an AI with unbounded ability to interact with the universe would most likely rapidly destroy all humans. It’s nonobvious that there is no possible way to code an AI that can reliably tell the difference between the two, and a solution to this problem naively seems rather more tractable than solving Friendliness in full generality. I’d like to see an exploration of this problem.
“When your gut won’t shut up and multiply.” The recent downvoted discussion post seems to be in this area, suggesting the wider community is perhaps less interested than I, but I’d love to see some practical advice on effective decision strategies when one’s calculated best action is intuitively morally dubious, with anecdotes of the success or failure of particular approaches.
“Times when I noticed I was confused”. In theory, noticing you’re confused sounds like an effective heuristic. But the explanation in the sequences only gave a retroactive example of when Eliezer should have applied it, and didn’t. I’d like to see more examples of when this has and hasn’t worked in practice, and useful habits to acquire that make you more likely to be able to notice.
“Times when I noticed I was confused”. In theory, noticing you’re confused sounds like an effective heuristic. But the explanation in the sequences only gave a retroactive example of when Eliezer should have applied it, and didn’t. I’d like to see more examples of when this has and hasn’t worked in practice, and useful habits to acquire that make you more likely to be able to notice.
Most of my examples here are trite individually, but significant collectively; that is, I remember the habit more easily than any particular examples. There have been situations where I had some niggling doubt, said “I’m confused, I ought to resolve this uncertainty,” and after research concluded that I was wrong and by acting early I saved myself some hardship. But while I’m certain there have been at least three of those, I have trouble remembering them or thinking that the ones I do remember are worth sharing.
Reflecting back on LessWrong’s past, I’ve noticed a pattern of article voting that seems almost striking to me: Questions do not get upvoted nearly on the same order as answers do.
Perhaps it would be useful to have a thread where LessWrong could posit topics and upvote the article titles that it would be most interested in reading? For example, I am now drafting a post titled “Applying Bayes Theorem.” Provided I can write high-quality content under that title, I expect LessWrong would be intensely interested in this on account of not fully grasping exactly how to do so.
So as a trial run: What topics currently elude your understanding, and what might the title of a high-quality article that addressed that topic be?
“Lower Bounds on Superintelligence”. While a lot of LW content is carefully researched, much of what’s posted in support of the singularity hypothesis seems to devolve into just-so stories. I’d like to see a dry, carefully footnoted argument for why an intelligence that was able to derive correct theories from evidence, or generate creative ideas, much faster than humans would necessarily rapidly acquire the ability to eliminate all human life. In particular I’m looking for historical analogies, cases where new discoveries with important practical implications were definitely delayed not just due to e.g. industrial capacity, but solely through human stupidity.
“Trading with entities that are smarter than you”. Given the ability of highly intelligent entities to predict the future better than you can, and deceive without outright lying, what kind of trades or bets is it wise to enter into with such entities? What kind of safeguards would you need to have in place?
“How to get a stupid person to let you out of a box”. Along with, I think, many people who’ve never done it, I find the results of the AI-box experiment highly implausible. I can’t even imagine a superintelligent persuading me to let it out, or, equivalently, I can’t imagine persuading even someone very stupid to let me out. I know the most successful AI players are keeping their strategies secret for reasons I don’t understand (if nothing else, it seems to imply those strategies are exceedingly fragile), but if there’s anyone who has a robust strategy that’s even partially effective I’d be very interested to see it.
“From printing results to destroying all humans”—to me this is the weakest part of the MIRI et al case, and I think most objections we see are variants on this theme. It’s obvious that an oracle-like AI would have to interact with the universe in some sense. It’s obvious that an AI with unbounded ability to interact with the universe would most likely rapidly destroy all humans. It’s nonobvious that there is no possible way to code an AI that can reliably tell the difference between the two, and a solution to this problem naively seems rather more tractable than solving Friendliness in full generality. I’d like to see an exploration of this problem.
“When your gut won’t shut up and multiply.” The recent downvoted discussion post seems to be in this area, suggesting the wider community is perhaps less interested than I, but I’d love to see some practical advice on effective decision strategies when one’s calculated best action is intuitively morally dubious, with anecdotes of the success or failure of particular approaches.
“Times when I noticed I was confused”. In theory, noticing you’re confused sounds like an effective heuristic. But the explanation in the sequences only gave a retroactive example of when Eliezer should have applied it, and didn’t. I’d like to see more examples of when this has and hasn’t worked in practice, and useful habits to acquire that make you more likely to be able to notice.
Most of my examples here are trite individually, but significant collectively; that is, I remember the habit more easily than any particular examples. There have been situations where I had some niggling doubt, said “I’m confused, I ought to resolve this uncertainty,” and after research concluded that I was wrong and by acting early I saved myself some hardship. But while I’m certain there have been at least three of those, I have trouble remembering them or thinking that the ones I do remember are worth sharing.
That’s the kind of position I see frequently here—but from the outside it’s very unconvincing. So I’d very much like to see concrete examples.