leogao

62

Very exciting that JumpReLU works well with STE gradient estimation! I think this fixes one of the biggest flaws with TopK, which is that having a fixed number of latents k on each token is kind of wonky. I also like the argument in section 4 a lot - in particular the point about how this works because we're optimizing the expectation of the loss. Because of how sparse the features are, I wonder if it would reduce gradient noise substantially to use a KDE with state persisting across a few recent steps.

60

Some takes on some of these research questions:

Looking for opposing feature directions in SAEs

I checked a top-k SAE with 256k features and k=256 trained on GPT-4 and found only 286 features that had any other feature with cosine similarity < -0.9, and 1314 with cosine sim < -0.7.

SAE/Transcoder activation shuffling

I'm confident that when learning rate and batch size are tuned properly, not shuffling eventually converges to the same thing as shuffling. The right way to frame this imo is the efficiency loss from not shuffling, which from preliminary experiments+intuition I'd guess is probably substantial.

How much does initializing the encoder to be the transpose of the decoder (as done so here and here) help for SAEs and transcoders?

It helps tremendously for SAEs by very substantially reducing dead latents; see appendix C.1 in our paper.

92

Cool work - figuring out how much of scaling up autoencoders is discovering new features vs splitting existing ones feels quite important. Especially since for any one scale of autoencoder there are simultaneously features which are split too finely and features which are too rare to yet be discovered, it seems quite plausible that the most useful autoencoders will be ones with features stitched together from multiple scales.

Some minor nitpicks: I would recommend always thinking of MSE/L0 in terms of the frontier between the two, rather than either alone; in my experiments I found it very easy to misjudge at a glance whether a run with better MSE but worse L0 was better or worse than the frontier.

50

VAEs:

a normal autoencoder decodes single latents z to single images (or whatever other kind of data) x, and also encodes single images x to single latents z.

with VAEs, we want our decoder (p(x|z)) to take single latents z and output a distribution over x's. for simplicity we generally declare that this distribution is a gaussian with identity covariance, and we have our decoder output a single x value that is the mean of the gaussian.

because each x can be produced by multiple z's, to run this backwards you also need a distribution of z's for each single x. we call the ideal encoder p(z|x) - the thing that would perfectly invert our decoder p(x|z). unfortunately, we obviously don't have access to this thing. so we have to train an encoder network q(z|x) to approximate it. to make our encoder output a distribution, we have it output a mean vector and a stddev vector for a gaussian. at runtime we sample a random vector eps ~ N(0, 1) and multiply it by the mean and stddev vectors to get an N(mu, std).

to train this thing, we would like to optimize the following loss function:

-log p(x) + KL(q(z|x)||p(z|x))

where the terms optimize the likelihood (how good is the VAE at modelling data, assuming we have access to the perfect z distribution) and the quality of our encoder (how good is our q(z|x) at approximating p(z|x)). unfortunately, neither term is tractable - the former requires marginalizing over z, which is intractable, and the latter requires p(z|x) which we also don't have access to. however, it turns out that the following is mathematically equivalent and is tractable:

-E z~q(z|x) [log p(x|z)] + KL(q(z|x)||p(z))

the former term is just the likelihood of the real data under the decoder distribution given z drawn from the encoder distribution (which happens to be exactly equivalent to the MSE, because it's the log of gaussian pdf). the latter term can be computed analytically, because both distributions are gaussians with known mean and std. (the distribution p is determined in part by the decoder p(x|z), but that doesn't pin down the entire distribution; we still have a degree of freedom in how we pick p(z). so we typically declare by fiat that p(z) is a N(0, 1) gaussian. then, p(z|x) is implied to be equal to p(x|z) p(z) / sum z' p(x|z') p(z'))

80

We had done very extensive ablations at small scale where we found TopK to be consistently better than all of the alternatives we iterated through, and by the time we launched the big run we had already worked out how to scale all of the relevant hyperparameters, so we were decently confident.

One reason we might want a progressive code is it would basically let you train one autoencoder and use it for any k you wanted to at test time (which is nice because we don't really know exactly how to set k for maximum interpretability yet). Unfortunately, this is somewhat worse than training for the specific k you want to use, so our recommendation for now is to train multiple autoencoders.

Also, even with a progressive code, the activations on the margin would not generally be negative (we actually apply a ReLU to make sure that the activations are definitely non-negative, but almost always the (k+1)th value is still substantially positive)

110

To add some more concreteness: suppose we open up the model and find that it's basically just a giant k nearest neighbors (it obviously can't be literally this, but this is easiest to describe as an analogy). Then this would explain why current alignment techniques work and dissolves some of the mystery of generalization. Then suppose we create AGI and we find that it does something very different internally that is more deeply entangled and we can't really make sense of it because it's too complicated. Then this would imo also provide strong evidence that we should expect our alignment techniques to break.

In other words, a load bearing assumption is that current models are fundamentally simple/modular in some sense that makes interpretability feasible, and that observing this breaking in the future is probably important evidence that will hopefully come before those future systems actually kill everyone.

90

Thanks for your kind words!

My views on interpretability are complicated by the fact that I think it's quite probable there will be a paradigm shift between current AI and the thing that is actually AGI like 10 years from now or whatever. So I'll describe first a rough sketch of what I think within-paradigm interp looks like and then what it might imply for 10 year later AGI. (All these numbers will be very low confidence and basically made up)

I think the autoencoder research agenda is currently making significant progress on item #1. The main research bottlenecks here are (a) SAEs might not be able to efficiently capture *every* kind of information we care about (e.g circular features) and (b) residual stream autoencoders are not exactly the right thing for finding circuits. Probably this stuff will take a year or two to really hammer out. Hopefully our paper helps here by giving a recipe to push autoencoders really quickly so we bump into the limitations faster and with less second guessing about autoencoder quality.

Hopefully #4 can be done to some great part in parallel with #1; there's a whole bunch of engineering needed to e.g take autoencoders and scale them up to capture all the behavior of the model (which was also a big part of the contribution of this paper). I'm pretty optimistic that if we have a recipe for #1 that we trust, the engineering (and efficiency improvements) for scaling up is doable. Maybe this adds another year of serial time. The big research uncertainty here fmpov is how hard it is to actually identify the structures we're looking for, because we'll probably have a tremendously large sparse network where each node does some really boring tiny thing.

However, I mostly expect that GPT-4 (and probably 5) is probably just actually not doing anything super spicy/stabby. So I think most of the value of doing this interpretability will be to sort of pull back the veil, so to speak, of how these models are doing all the impressive stuff. Some theories of impact:

- Maybe we'll become less confused about the nature of intelligence in a way that makes us just have better takes about alignment (e.g there will be many mechanistic theories of what the heck GPT-4 is doing that will have been conclusively ruled out)
- Maybe once the paradigm shift happens, we will be better prepared to identify exactly what interpretability assumptions it broke (or even just notice whether some change is causing a mechanistic paradigm shift)

Unclear what timeline these later things happen on; probably depends a lot on when the paradigm shift(s) happen.

50

For what it's worth, it seems much more likely to me for catastrophic Goodhart to happen because the noise isn't independent from the thing we care about, rather than the noise being independent but heavy tailed.

This is likely not the first instance, but OpenAI was already using the word "aligned" in this way in 2021 in the Codex paper.

https://arxiv.org/abs/2107.03374 (section 7.2)