LessWrong 2.0 Reader

One confusion I have with MAD as an approach to ELK is that it seems to assume some kind of initial inner alignment. If we're flagging when the model takes actions / makes predictions for "unusual reasons", where unusual is define with respect to some trusted set, but aligned and misaligned models are behaviorally indistinguishable on the trusted set, then a model could learn to do things for misaligned reasons on the trusted set, and then use those same reasons on the untrusted set. For example, a deceptively aligned model would appear aligned in training but attempt take-over in deployment for the "same reason" (e.g. to maximize paperclips), but a MAD approach that "properly" handles out of distribution cases would not flag take over attempts because we want models to be able to respond to novel situations. 

I guess this is part of what motivates measurement tampering as a subclass of ELK - instead of trying to track motivations of the agent as reasons, we try to track the reasons for the measurement predictions, and we have some trusted set with no tampering, where we know the reasons for the measurements is ~exactly that the thing we want to be measuring. 

Now time to check my answer by rereading https://www.alignmentforum.org/posts/vwt3wKXWaCvqZyF74/mechanistic-anomaly-detection-and-elk [AF · GW]

That’s a great idea, Thank you!

And here it is: https://manifold.markets/keltan/will-there-be-a-lessonline-2025

These days companies frequently buy back shares because they can do that without having to pay taxes for that. 

This makes a lot of sense actually.

If an investor wants to invest in a couple of different markets then that investor can choose some companies in those markets and buy shares. A single company in two non-synergistic markets is just silly from this perspective, if it was instead two companies investors would be more able to specifically choose which part they like.

If the company doesn't have shares (privately held) then this incentive does not exist.

In "Capitalism and Freedom" Milton Freidman has a section about diverse companies. The book is probably dated but some aspect of the 1960's American tax system meant that if a company you owned shares in paid you a dividend you paid tax on it, but if they re-invested a profit that was not taxed. This meant that, if (cartoon example) 100% of the shareholders in Company A wanted to take their dividends and invest them in a start-up that does X, the same thing can be done more tax efficiently by Company A instead not paying any dividends and using the money to set up a new arm that does X (the new arm essentially being a separate startup company in all but name).

Freidman thought this was a reasonably serious distortion on markets and that investors should have to pay tax on re-invested profits at the same rate as dividends, to correct for it.

Perhaps some aspect of tax systems in some countries is having a similar effect, that taking some money out of my lawnmower company to invest in a chip foundry would incur more taxes than expanding the company to do both lawnmowers and computer chips.

Is there a prediction market for that?

I don't think there is, but you could make one!

Our overall best guess is that an important role of early MLPs is to act as a “multi-token embedding”, that selects^[1] the right unit of analysis from the most recent few tokens (e.g. a name) and converts this to a representation (i.e. some useful meaning encoded in an activation). We can recover different attributes of that unit (e.g. sport played) by taking linear projections, i.e. there are linear representations of attributes. Though we can’t rule it out, our guess is that there isn’t much more interpretable structure (e.g. sparsity or meaningful intermediate representations) to find in the internal mechanisms/parameters of these layers. For future mech interp work we think it likely suffices to focus on understanding how these attributes are represented in these multi-token embeddings (i.e. early-mid residual streams on a multi-token entity), using tools like probing and sparse autoencoders, and thinking of early MLPs similar to how we think of the token embeddings, where the embeddings produced may have structure (e.g. a “has space” or “positive sentiment” feature), but the internal mechanism is just a look-up table with no structure to interpret.

You may be interested in works like REMEDI and Identifying Linear Relational Concepts in Large Language Models.

In Fig 1, is the vertical axis P(world) ?

It seems to me like many of those business conglomerates are privately held. The stock market seems to prefer if businesses parts that have no synergy with the main business get sold off. 

On the stock market it's easy for capital owners to diversify their assets. When companies are privately held building such business conglomerates might be the way to diversify.

One feature of polyamory is that it means continuous auditions of potential replacements by all parties. You are not trading up in the sense that you can have multiple partners, but one thing leads to another and there are only so many hours in the day.

Polyamory is not that different from monogamy in this respect. It's just that in monogamy "having a relationship" is a binary: either you have it or you don't have it. In polyamory, there is a scale, starting from "meeting once in a blue moon" all the way to "living together with kids and joint finances". So, if in monogamy your attitude might be "I will not trade up unless I meet someone x% better", then in polyamory your attitude might be "I will devote you y% of my time and will not reduce this number unless there's someone x% better competing for this slot". (And in both cases x might be very high.)

More generally, I feel that a lot of arguments against polyamory fail the "replace with platonic friendship" test. Like, monogamous people also have to somehow balance the time they invest in their relationship vs. friends vs. family vs. hobbies etc, and also have to balance the time allocated to different friends. I know that some mono people feel that sex is some kind of magic pixie dust which makes a relationship completely different and not comparable in any way to platonic friendship, but... Not everyone feels this way? (In both directions: I simultaneously consider romantic relationship comparable to "mere" platonic friendships and also consider platonic friendships substantially more important/committing than seems to be the culturally-prescribed attitude.)

Also, it feels like this discussion has a missing mood and/or a typical mind fallacy. For me, monogamy was a miserable experience. Even aside from the fact you only get to have one relationship, there's all the weird rules about which things are "inappropriate" (see survey in the OP) and also the need to pretend that you're not attracted to other people (Not All Mono, but I think many relationships are like that). All the "pragmatic" arguments about why polyamory is bad sound to me similar to hypothetical arguments that gay relationships are bad. I mean, there might be some aspects of gay relationships that are often worse than corresponding aspects of straight relationships. But if you're gay, a gay relationship is still way better for you! Even if you're bi and in some sense "have a choice", it still seems inappropriate to try convincing you about how hetero is much better.

Warning: About to get a little ranty/emotional, sorry about that but was hard to express otherwise.

Finally, not to be that girl, but it's a little insensitive to talk about this without the least acknowledgement that polyamory is widely stigmatized and discriminated against. I know it's LessWrong here, we're supposed to use decoupling norms and not contextualizing norms, and I'm usually fully in favor of that, but it still seems to me that this post would better on the margin, if it had a little in the way of acknowledging this asymmetry in the debate. 

Instead, the OP talks about "encouraging widespread adaptation". What?? I honestly don't know, maybe in the Mythic Bay Area, someone is encouraging widespread conversion to polyamory. In the rest of the world, we only want (i) not be stigmatized (ii) not be discriminated against (iii) having some minimal awareness that polyamory is even an option (it was certainly an eye-opening discovery for me!) and (iv) otherwise, being left alone, and not have mono people endlessly explain to us how their way is so much better [My spouse tells me this last bit was too combative. Sorry about that: we are certainly allowed to have respectful discussion about the comparative advantages of different lifestyles.]

Let's say: if you train a coherently goal-directed, situationally aware, somewhat-better-than-human-level model using baseline forms of self-supervised pre-training + RLHF on diverse, long-horizon, real-world tasks, my subjective probability is ~25% that this model will be performing well in training in substantial part as part of an instrumental strategy for seeking power for itself and/or other AIs later.

Have you tried extending this gut estimate to something like:

If many labs use somewhat different training procedures to train their models but that each falls under the umbrella of "coherently goal-directed, situationally aware [...]", what is the probability that at least one of these models "will be performing well in training in substantial part as part of an instrumental strategy for seeking power for itself and/or other AIs later."?