Wei Dai60

Traditionally, those techniques are focused on what the model is outputting, not what the model’s underlying motivations are. But I haven’t read all the literature. Am I missing something?

It's confusing to me as well, perhaps because different people (or even the same person at different times) emphasize different things within the same approach, but here's one post where someone said, "It is important that the overseer both knows which action the distilled AI wants to take as well as why it takes that action."

Wei Dai40

I'm confused about how heterogeneity in data quality interacts with scaling. Surely training a LM on scientific papers would give different results from training it on web spam, but data quality is not an input to the scaling law... This makes me wonder whether your proposed forecasting method might have some kind of blind spot in this regard, for example failing to take into account that AI labs have probably already fed all the scientific papers they can into their training processes. If future LMs train on additional data that have little to do with science, could that keep reducing overall cross-entropy loss (as scientific papers become a smaller fraction of the overall corpus) but fail to increase scientific ability?

Wei Dai42

Does anyone know why security amplification and meta-execution are rarely talked about these days? I did a search on LW and found just 1 passing reference to either phrase in the last 3 years. Is the problem not considered an important problem anymore? The problem is too hard and no one has new ideas? There are too many alignment problems/approaches to work on and not enough researchers?

Wei Dai30

I gave this explanation at the start of the UDT1.1 post:

When describing UDT1 solutions to various sample problems, I've often talked about UDT1 finding the function S* that would optimize its preferences over the world program P, and then return what S* would return, given its input. But in my original description of UDT1, I never explicitly mentioned optimizing S as a whole, but instead specified UDT1 as, upon receiving input X, finding the optimal output Y* for that input, by considering the logical consequences of choosing various possible outputs. I have been implicitly assuming that the former (optimization of the global strategy) would somehow fall out of the latter (optimization of the local action) without having to be explicitly specified, due to how UDT1 takes into account logical correlations between different instances of itself. But recently I found an apparent counter-example to this assumption.

Wei Dai31

Then it would repeat the same process for t=1 and the copy. Conditioned on “I will see C” at t=1, it will conclude “I will see CO” with probability 1⁄2 by the same reasoning as above. So overall, it will assign:p(“I will see OO”) = 1⁄2,p(“I will see CO”) = 1⁄4,p(“I will see CC”) = 1⁄4

  1. If we look at the situation in 0P, the three versions of you at time 2 all seem equally real and equally you, yet in 1P you weigh the experiences of the future original twice as much as each of the copies.
  2. Suppose we change the setup slightly so that copying of the copy is done at time 1 instead of time 2. And at time 1 we show O to the original and C to the two copies, then at time 2 we show them OO, CO, CC like before. With this modified setup, your logic would conclude P(“I will see O”)=P(“I will see OO”)=P(“I will see CO”)=P(“I will see CC”)=1/3 and P(“I will see C”)=2/3. Right?
  3. Similarly, if we change the setup from the original so that no observation is made at time 1, the probabilities also become P(“I will see OO”)=P(“I will see CO”)=P(“I will see CC”)=1/3.
  4. Suppose we change the setup from the original so that at time 1, we make 999 copies of you instead of just 1 and show them all C before deleting all but 1 of the copies. Then your logic would imply P("I will see C")=.999 and therefore P(“I will see CO”)=P(“I will see CC”)=0.4995, and P(“I will see O”)=P(“I will see OO”)=.001.

This all make me think there's something wrong with the 1/2,1/4,1/4 answer and with the way you define probabilities of future experiences. More specifically, suppose OO wasn't just two letters but an unpleasant experience, and CO and CC are both pleasant experiences, so you prefer "I will experience CO/CC" to "I will experience OO". Then at time 0 you would be willing to pay to switch from the original setup to (2) or (3), and pay even more to switch to (4). But that seems pretty counterintuitive, i.e., why are you paying to avoid making observations in (3), or paying to make and delete copies of yourself in (4). Both of these seem at best pointless in 0P.

But every other approach I've seen or thought of also has problems, so maybe we shouldn't dismiss this one too easily based on these issues. I would be interested to see you work out everything more formally and address the above objections (to the extent possible).

Wei Dai40

Defining the semantics and probabilities of anticipation seems to be a hard problem. You can see some past discussions of the difficulties at The Anthropic Trilemma and its back-references (posts that link to it). (I didn't link to this earlier in case you already found a fresh approach that solved the problem. You may also want to consider not reading the previous discussions to avoid possibly falling into the same ruts.)

Wei Dai60

We can assign meanings to statements like “my sensor sees red” by picking out subsets of experiences, just as before.

How do you assign meaning to statements like "my sensor will see red"? (In the OP you mention "my sensors will see the heads side of the coin" but I'm not sure what your proposed semantics of such statements are in general.)

Also, here's an old puzzle of mine that I wonder if your line of thinking can help with: At time 1 you will be copied and the original will be shown "O" and the copy will be shown "C", then at time 2 the copy will be copied again, and the three of you will be shown "OO" (original), "CO" (original of copy), "CC" (copy of copy) respectively. At time 0, what are your probabilities for "I will see X" for each of the five possible values of X?

Wei Dai41

Yeah, this makes sense, thanks. I think I've read one or maybe both of your posts, which is probably why I started having second thoughts about my comment soon after posting it. :)

Wei Dai67

I agree with this point as stated, but think the probability is more like 5% than 0.1%.

How do you define or think about "light" in "light RLHF" when you make a statement like this, and how do you know that you're thinking about it the same way that Alex is? Is it a term of art that I don't know about, or has it been defined in a previous post?

Wei Dai20

In any case, I'd be interested in an explanation of what the differences are and why one type of training produces schemers and the other is very unlikely to.

Thinking over this question myself, I think I've found a reasonable answer. Still interested in your thoughts but I'll write down mine:

It seems like evolution "wanted" us to be (in part) reward-correlate maximizers (i.e., being a reward-correlate maximizer was adaptive in our ancestral environment), and "implemented" this by having our brains internally do "heavy RL" throughout our life. So we become reward-correlate maximizing agents early in life, and then when our parents do something like RLHF on top of that, we become schemers pretty easily.

So the important difference is that with "pretraining + light RLHF" there's no "heavy RL" step.

Load More