LessWrong 2.0 Reader

Going by today's standards, we should have banned Gwern in 2012.

And I think that would have been a mistake.

I wonder how many other mistakes we made. The problem is, we won't get good feedback on this.

Love this!!! To inspect the concept of neutrality as a software like process driven institution is a very illuminating approach. 

I would add that there is a very relevant discussion in the AI sphere which is profoundly connected to the points you are making: the existence of bias in AI models. 

To identify and measure bias is in a sense to identify and measure lack of neutrality, so it follows that, to define bias, one must first be very rigorous on the definition of neutrality. 

This can seem simple for some of the more pedestrian AI tasks, but can become increasingly sophisticated as we introduce AI as an essential piece in workflows and institutions. 

AI algorithms can be heavily biased, datasets can be biased and even data structures can be biased. 

I feel this is a topic which you can further explore in the future. Thank you for this!

I'm curious about Claude's coauthorship of this post. How did it help?

The crucial point is: when the untrusted model gets to step 6, it sees R in the history instead of A, so it knows its attempt was replaced.

Why does it know its attempt was replaced? Because it can guess it wouldn't have produced R? If it can tell that, how come it can't guess it specifically would have produced A?

also, this 'subconscious parts going on strike' theory makes slightly different predictions than the 'is it good for the whole system/live' theory

for instance, i predict that you can have 'dead parts' that e.g. give people social anxiety based on past trauma, even though it's no longer actually relevant to their current situation.

and that if you override this social anxiety using 'live willpower' for a while, you can get burnout, even though the willpower is in some sense 'correct' about what would be good for the overall flourishing of the system given the current reality.

I think what Justin is saying is that finding a single monogamous partner is not significantly different from finding two, three, etc. For some things you only care about succeeding once. So a 63% chance of success (any number of times) means a .63 expected value (because all successes after the first have a value of 0).

Meanwhile for other things, such as polyamorous partners, 2 partners is meaningfully better than one, so the expected value truly is 1, because you will get one partner on average. (Though this assumes 2 partners is twice as good as one, we can complicate this even more if we assume that 2 partners is better, but not twice as good)

In case it's a helpful data point: lines of reasoning sorta similar to the ones around the infohazard warning seemed to have interesting and intense psychological effects on me one time. It's hard to separate out from other factors, though, and I think it had something to do with the fact that lately I've been spending a lot of time learning to take ideas seriously on an emotional level instead of only an abstract one.

If you look at the Active Inference community there's a lot of work going into PPL-based languages to do more efficient world modelling but that shit ain't easy and as you say it is a lot more compute heavy.

I think there'll be a scaling break due to this but when it is algorithmically figured out again we will be back and back with a vengeance as I think most safety challenges have a self vs environment model as a necessary condition to be properly engaged. (which currently isn't engaged with LLMs wolrd modelling)

There was one comment on twitter that the RLHF-finetuned models also still have the ability to play chess pretty well, just their input/output-formatting made it impossible for them to access this ability (or something along these lines). But apparently it can be recovered with a little finetuning. 

Indeed. I fixed it. Let's see whether it repeats itself (we got kind of malformed HTML from the RSS feed).