3

This is a special post for quick takes by leogao. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
Mentioned in
[-]leogao171

learning thread for taking notes on things as i learn them (in public so hopefully other people can get value out of it)

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'))

One possible model of AI development is as follows: there exists some threshold beyond which capabilities are powerful enough to cause an x-risk, and such that we need alignment progress to be at the level needed to align that system before it comes into existence. I find it informative to think of this as a race where for capabilities the finish line is x-risk-capable AGI, and for alignment this is the ability to align x-risk-capable AGI. In this model, it is necessary but not sufficient for alignment for alignment to be ahead by the time it's at the finish line for good outcomes: if alignment doesn't make it there first, then we automatically lose, but even if it does, if alignment doesn't continue to improve proportional to capabilities, we might also fail at some later point. However, I think it's plausible we're not even on track for the necessary condition, so I'll focus on that within this post.

Given my distributions over how difficult AGI and alignment respectively are, and the amount of effort brought to bear on each of these problems, I think there's a worryingly large chance that we just won't have the alignment progress needed at the critical juncture.

I also think it's plausible that at some point before when x-risks are possible, capabilities will advance to the point that the majority of AI research will be done by AI systems. The worry is that after this point, both capabilities and alignment will be similarly benefitted by automation, and if alignment is behind at the point when this happens, then this lag will be "locked in" because an asymmetric benefit to alignment research is needed to overtake capabilities if capabilities is already ahead.

There are a number of areas where this model could be violated:

• Capabilities could turn out to be less accelerated than alignment by AI assistance. It seems like capabilities is mostly just throwing more hardware at the problem and scaling up, whereas alignment is much more conceptually oriented.
• After research is mostly/fully automated, orgs could simply allocate more auto-research time to alignment than AGI.
• Alignment(/coordination to slow down) could turn out to be easy. It could turn out that applying the same amount of effort to alignment and AGI results in alignment being solved first.

However, I don't think these violations are likely for the following reasons respective:

• It's plausible that our current reliance on scaling is a product of our theory not being good enough and that it's already possible to build AGI with current hardware if you have the textbook from the future. Even if the strong version of the claim isn't true, one big reason that the bitter lesson is true is that bespoke engineering is currently expensive, and if it became suddenly a lot cheaper we would see a lot more of it and consequently squeezing more out of the same hardware. It also seems likely that before total automation, there will be a number of years where automation is best modelled as a multiplicative factor on human researcher effectiveness. In that case, because of the sheer number of capabilities researchers compared to alignment researchers, alignment researchers would have to benefit a lot more to just break even.
• If it were the case that orgs would pivot, I would expect them to currently be allocating a lot more to alignment than they do currently. While it's still plausible that orgs haven't allocated more to alignment because they think AGI is far away, and that a world where automated research is a thing is a world where orgs would suddenly realize how close AGI is and pivot, that hypothesis hasn't been very predictive so far. Further, because I expect the tech for research automation to be developed at roughly the same time by many different orgs, it seems like not only does one org have to prioritize alignment, but actually a majority weighted by auto research capacity have to prioritize alignment. To me, this seems difficult, although more tractable than the other alignment coordination problem, because there's less of a unilateralist problem. The unilateralist problem still exists to some extent: orgs which prioritize alignment are inherently at a disadvantage compared to orgs that don't, because capabilities progress feeds recursively into faster progress whereas alignment progress is less effective at making future alignment progress faster. However, on the relevant timescales this may become less important.
• I think alignment is a very difficult problem, and that moreover by its nature it's incredibly easy to underestimate. I should probably write a full post about my take on this at some point, and I don't really have space here to really dive into it here, but a quick meta level argument for why we shouldn't lean on alignment easiness even if there is a non negligible chance of easiness is that a) given the stakes, we should exercise extreme caution and b) there are very few problems we have that are in the same reference class as alignment, and of the few that are even close, like computer security, they don't inspire a lot of confidence.

I think exploring the potential model violations further is a fruitful direction. I don't think I'm very confident about this model.

one man's modus tollens is another man's modus ponens:

"making progress without empirical feedback loops is really hard, so we should get feedback loops where possible" "in some cases (i.e close to x-risk), building feedback loops is not possible, so we need to figure out how to make progress without empirical feedback loops. this is (part of) why alignment is hard"

Yeah something in this space seems like a central crux to me.

I personally think (as a person generally in the MIRI-ish camp of "most attempts at empirical work are flawed/confused"), that it's not crazy to look at the situation and say "okay, but, theoretical progress seems even more flawed/confused, we just need to figure out some how of getting empirical feedback loops."

I think there are some constraints on how the empirical work can possibly work. (I don't think I have a short thing I could write here, I have a vague hope of writing up a longer post on "what I think needs to be true, for empirical work to be helping rather than confusedly not-really-helping")

A few axes along which to classify optimizers:

• Competence: An optimizer is more competent if it achieves the objective more frequently on distribution
• Capabilities Robustness: An optimizer is more capabilities robust if it can handle a broader range of OOD world states (and thus possible pertubations) competently.
• Generality: An optimizer is more general if it can represent and achieve a broader range of different objectives
• Real-world objectives: whether the optimizer is capable of having objectives about things in the real world.

Some observations: it feels like capabilities robustness is one of the big things that makes deception dangerous, because it means that the model can figure out plans that you never intended for it to learn (something not very capabilities robust would just never learn how to deceive if you don't show it). This feels like the critical controller/search-process difference: controller generalization across states is dependent on the generalization abilities of the model architecture, whereas search processes let you think about the particular state you find yourself in. The actions that lead to deception are extremely OOD, and a controller would have a hard time executing the strategy reliably without first having seen it, unless NN generalization is wildly better than I'm anticipating.

Real world objectives is definitely another big chunk of deception danger; caring about the real world leads to nonmyopic behavior (though maybe we're worried about other causes of nonmyopia too? not sure tbh), I'm actually not sure how I feel about generality: on the one hand, it feels intuitive that systems that are only able to represent one objective have got to be in some sense less able to become more powerful just by thinking more; on the other hand I don't know what a rigorous argument for this would look like. I think the intuition relates to the idea of general reasoning machinery being the same across lots of tasks, and this machinery being necessary to do better by thinking harder, and so any model without this machinery must be weaker in some sense. I think this feeds into capabilities robustness (or lack thereof) too.

Examples of where things fall on these axes:

• A rock would be none of the properties.
• A pure controller (i.e a thermostat, "pile of heuristics") can be competent, but not as capabilities robust, not general at all, and have objectives over the real world.
• An analytic equation solver would be perfectly competent and capablilities robust (if it always works), not very general (it can only solve equations), and not be capable of having real world objectives.
• A search based process can be competent, would be more capabilities robust and general, and may have objectives over the real world.
• A deceptive optimizer is competent, capabilities robust, and definitely has real world objectives