LessWrong 2.0 Reader

Let D be the distribution of (reward, trajectory) pairs for every possible trajectory.

Split D into two subsets: D1 where reward > 7 and D2 where reward ≤ 7.

Suppose that, in D, 1-in-a-googol samples is in the subset D1, and all the rest are in D2.

Now we randomly pick a million samples from D in an attempt to learn the distribution D. But (as expected with overwhelming probability), it turns out that every single one of those million samples are more specifically in the subset D2.

Now consider a point X in D1 (its reward is 30). Question: Is X “out of distribution”?

Arguably no, because we set up a procedure to learn the distribution D, and D contains X.

Arguably yes, because when we ran the procedure, all the points actually used to learn the distribution were in D2, so we were kinda really only learning D2, and D2 doesn’t contain X.

Anyway, I was gonna say the opposite of what you said—sufficiently hard optimization via conditional sampling works in theory (i.e., if you could somehow learn D and conditionally sample on reward>7, it would work), but not in practice (because reward>7 is so hard to come by that you will never actually learn that part of D by random sampling).

I guess my question would be 'how else did you think a well-generalising sequence model would achieve this?' Like, what is a sufficient world model but a posterior over HMM states in this case? This is what GR theorem asks. (Of course, a poorly-fit model might track extraneous detail or have a bad posterior.)

From your preamble and your experiment design, it looks like you correctly anticipated the result, so this should not have been a surprise (to you). In general I object to being sold something as surprising which isn't (it strikes me as a lesser-noticed and perhaps oft-inadvertent rhetorical dark art and I see it on the rise on LW, which is sad).

That said, since I'm the only one objecting here, you appear to be more right about the surprisingness of this!

The linear probe is new news (but not surprising?) on top of GR, I agree. But the OP presents the other aspects as the surprises, and not this.

Note that I agree with your sentiment here, although my concrete argument is basically what LawrenceC wrote as a reply to this post [LW(p) · GW(p)].

Ryan, this is kind of a side-note but I notice that you have a very Paul-like approach to arguments and replies on LW.

Two things that come to notice:

1. You have a tendency to reply to certain posts or comments with "I don't quite understand what is being said here, and I disagree with it." or, "It doesn't track with my views", or equivalent replies that seem not very useful for understanding your object level arguments. (Although I notice that in the recent comments I see, you usually postfix it with some elaboration on your model.)
2. In the comment I'm replying to, you use a strategy of black-box-like abstraction modeling of a situation to try to argue for a conclusion, one that usually involves numbers such as multipliers or percentages. (I have the impression that Paul uses this a lot, and one concrete example that comes to mind is the takeoff speeds essay. I usually consider such arguments invalid when they seem to throw away information we already have, or seem to use a set of abstractions that don't particularly feel appropriate to the information I believe we have.

I just found this interesting and plausible enough to highlight to you. Its a moderate investment of my time to find out examples from your comment history to highlight all these instances, but writing this comment still seemed valuable.

Relevant: My Taxonomy of AI-risk counterarguments [LW · GW], inspired by Zvi Mowshowitz's The Crux List [LW · GW].

This is a really well-written response. I'm pretty impressed by it.

Thanks for posting Mako. I even mention Effective Altruism/Longtermism at one point in the video!

If your acceptable lower limit for basically anything is zero you wont be allowed to do anything, really anything. You have to name some quantity of capabilities progress that’s okay to do before you’ll be allowed to talk about AI in a group setting.

"The optimal amount of fraud is non-zero."

Looks relevant to me on a skim! I'd probably want to see some arguments in the submission for why this is useful tooling for mech interp people specifically (though being useful to non mech interp people too is a bonus!)

Originally I felt happy about these, because “mostly agreeing” is an unusually positive outcome for that opening. But these discussions are grueling. It is hard to express kindness and curiosity towards someone yelling at you for a position you explicitly disclaimed. Any one of these stories would be a success but en masse they amount to a huge tax on saying anything about veganism, which is already quite labor intensive.

The discussions could still be worth it if it changed the arguer’s mind, or at least how they approached the next argument. But I don’t get the sense that’s what happens. Neither of us have changed our minds about anything, and I think they’re just as likely to start a similar fight the next week.

May I ask what your motivation was when you decided to spend your time on the aforementioned discussions? Were you hoping to learn something or to persuade the other, or both?

It sounds to me as though the solution here is to be more cautious before spending time arguing. Sometimes (often) it is IMO wiser to cut your losses and stop replying.