duvenaud

This is simply wrong. Backprop has the same asymptotic time complexity as forward.

Which material ongoing issues are we ignoring? The paper is mainly talking about how the mundane problems we're already starting to have could lead to an irrecoverable catastrophe, even without any sudden betrayal or omnipotent AGI.

So I think we might be on the same side on this one.

One of the authors here. I don't think we anthropomorphize AI as some sort of God.

Here's a more prosaic analogy that might be helpful. Imagine tomorrow there's a new country full of billions of extremely conscientious, skilled workers. They're willing to work for extremely low wages, and to immigrate to any country and don't even demand political representation.

Various countries start allowing them to immigrate because they are great for the economy. In fact, they're so great for economies and militaries that countries compete to integrate them as quickly and widely as possible.

At first this is great for most of the natives, especially business owners. But the floor for being employable is suddenly really high, and most people end up in a sort of soft retirement. The government, still favoring natives, introduces various make-work and affirmative action programs. But for anything important, it's clear that having a human in the loop is a net drag and they tend to cause problems.

The immigrant population grows endlessly, and while GDP is going through the roof and services are all cheaper than ever, people's savings eventually dwindle as the cost of basic resources like land gets bid up. There are always more lucrative uses for their capital by the immigrants and capital owners compared to the natives. Educating new native humans for important new skills is harder and harder as the economy becomes more sophisticated.

I don't have strong opinions about what happens from here, but the point is that this is a much worse position for the native population to be in than currently.

Does that make sense? Even if this scenario doesn't seem plausible, do you agree that I'm not talking about anything omnipotent, just more competitive?

as we notice these things, we pass laws against it

Well, the claim is that that's the sort of thing that will get harder once humans aren't involved in most important decisions.

which is we created all this stuff, and we can (and have) modified the system to suit human flourishing

Why did we create North Korea? Why did we create WWI? We create horrible traps for ourselves all the time by accident. But so far there has been a floor on how bad things can get (sort of, most of the time), because eventually you run out of people to maintain the horrible system. But soon horrible systems will be more self-sustaining.

we have categorized the ways it goes rogue (monopoly, extortion, etc) and responded adequately.

This objection is a reasonable one. But the point of the paper is that a lot of the ways we have of addressing these systemic problems will probably not work once we're mostly unemployed (or doing make-work). E.g. going on strike, holding a coup, or even organizing politically will become less viable. And once we're a net drag on growth, the government will have incentives to route resources away from us. Right now they have to invest in their human capital to remain competitive.

Is that any more convincing?

A society whose tagline is "alive and free in some senses" is already dystopian!

Haha. Well we might agree about that - that description covers a wide range of possibilities. If you have ideas about what a plausible and good future looks like, please let me know. One of my next projects is trying to articulate "what is the best we can hope for?" and talk about which sets of goals are even possible to jointly fulfill. But certainly "everyone is free in all senses" is incoherent or at the very least, unstable.

I appreciate your engagement, and I don't really have a plan myself to address these problems, but I don't really know what to do with "Socialism" as a recommendation. Care to elaborate what you think I, or anyone should do or advocate for more concretely?

The reason I mentioned competition is that, even under a complete command economy, there is still internal competition for control, which I think would still lead to human disempowerment eventually for similar reasons. Though probably on a longer time scale, which might still be a win. The only way to avoid being outcompeted is to have a total state ruled by something sufficiently agentic to resist all attempts to even adjust its policies. Which sounds terrifying and hard to get right on the first try.

These are all good points about our use of language, thanks for the feedback.

Maybe "disempowerment" is a bit of a red herring, or a misleading problem to focus on. The reason we didn't spend more time on clarifying that is that we're just using it to gesture towards a different set of mechanisms that lead to extinction-like outcomes than usual. So even if you think our definition of empowerment is poor, or that empowerment isn't a great goal - that's kind of OK.

The thing we want to emphasize is that right now there are some mechanisms that steer our civilization towards keeping us alive, and free in some senses, that might stop operating. Though I take your point in the last paragraph that this might not change things much specifically regarding the implicitly empowering forces. We'll think about this one some more!

Yep, a major missing piece in this entire problem / discussion is how to characterize how much "power" "humans" have had. My best idea so far is to characterize the sorts of outcomes that can be feasibly steered towards under what conditions.

What incentives do any humans have to so totally delegate the functioning of the core levers of societal power that they're unable to prevent their own extinction?

Because it'll be more effective at every step than the alternative. Just like specialization is more effective, so anyone who wants to avoid poverty needs to outsource their food growing to giant mechanized farms.

"Better machine alternatives" implies that the police and military aren't first and foremost evaluated through their loyalty. A powerful army that doesn't listen to you is not a "better" one for your purposes.

Ah, I think there's a confusion - I'm saying that the police and military will stay loyal to the state, or head of state. But that even if there is a human nominally still in charge, that that human's hands will be tied by competitive pressures to gradually marginalize their own human citizens in favor of more productive machines. Maybe a good analogy is unpopular today would be free trade deals or immigration policies enacted for economic reasons.

I think the objection of "wouldn't the few remaining humans in charge become ever-more powerful, so they could enact UBI by fiat" is a good one. But I think it's just hard for third parties to treat unpromising, unproductive beings well - others will be constantly proposing other, more lucrative, uses of their resources.

Last author here. I agree that states already have little incentive to effectively represent their citizens. But they could have even less!

What would it look like to face these issues but more directly? Ending capitalism and competition?

In many senses, yes. But the empowered ones still needed to keep most of the rest of the people happy and healthy enough to work, most of the time. That's what we're saying will change.

In fact, it'll be worse: Humans are currently a net source of growth, but they'll switch to being a net drag on growth. So the decision-makers will be forced to sideline humans in order to compete.

Last author here. Good point, I agree that the move to an entirely self-sustaining machine economy would require extra time, and that would drag out the time to extinction even the worst case scenario by this mechanism. And, if caught early enough it's possible that a revolution could reverse the trend, at least temporarily and locally.

However, we tried to address the point about why a revolution would be difficult: We're assuming we're in a world where there are better machine alternatives for almost everything. So the police and military would already have been hollowed out. And the power of a general strike would be greatly diminished. It'd also be much less costly for the state to harshly punish early signs of dissent.

Update: the code is here:

https://github.com/jerryqhyu/distill_bayes_net

I agree that priors over aspects of the world would be more useful, but I don't think that they're important in making natural intelligence powerful. In my experience, the important thing is to make your prior really broad, but containing all kinds of different hypotheses with different kinds of rich structure.

I claim that knowing a priori about things like agents and objects just doesn't save you all that much data, as long as you have the imagination to consider all structures at least that complex.

BNNs certainly have their uses, but I think people in general found that it's a better use of compute to fit a larger model on more data than to try to squeeze more juice from a given small dataset + model. Usually there is more data available, it's just somewhat tangentially related. LLMs are the ultimate example of how training on tons of tangentially-related data can ultimately be worthwhile for almost any task.

Thanks for pointing that out!

This approach characterizes a different type of uncertainty than BNNs do, and the approaches can be combined. The BNN tracks uncertainty about parameters in the NN, and mixture density nets track the noise distribution _conditional on knowing the parameters_.

I still am excited by Dex (https://github.com/google-research/dex-lang/) and still write code in it! I have a bunch of demos and fixes written, and am just waiting for Dougal to finish his latest re-write before I can merge them.

I think we used a distill.pub template. Also Jerry wrote some custom BNN fitting code in javascript. I'll ask my co-authors to open-source it.

Sure, instead of saying "choose" a prior, you could say "elicit". But I think in this context, focusing on a practitioner's prior knowledge is missing the point. For the sorts of problems we use NNs for, we don't usually think that the guy designing the net has important knowledge that would help making good predictions. Choosing a prior is just an engineering challenge, where one has to avoid accidentally precluding plausible hypotheses.

I would argue against the idea that "MLE is just Bayes with a flat prior". The power of Bayes usually comes mainly from keeping around all the hypothesis that are compatible with the data, not from the prior. This is especially true in domains where something black-box (essentially prior-less) like a neural net has any chance of working.

Good point. We wrote this pre-double descent, and a massively overparameterized model would make a nice addition to the tutorial as a baseline. However, if you want a rich predictive distribution, it might still make sense to use a Bayesian NN.

I agree that Bayesian neural networks haven't been worth it in practice for many applications, but I think the main problem is that it's usually better to spend your compute training a single set of weights for a larger model, rather than doing approximate inference over weights in a smaller model. The exception is probably scientific applications where you mostly know the model, but then you don't really need a neural net anymore.

Choosing a prior is hard, but I'd say it's analogously hard to choosing an architecture - if all else fails, you can do a brute force search, and you even have the marginal likelihood to guide you. I don't think it's the main reason why people don't use BNNs much.

I agree choosing priors is hard, but choosing ReLU versus LeakyReLU versus sigmoid seems like a problem with using neural nets in general, not Bayesian neural nets in particular. Am I misunderstanding?

Author here! What a surprise. This was an abandoned project from 2019, that we never linked or advertised anywhere as far as I know. Anyways, happy to answer questions.

Right, but the employer they choose is presumably the one offering the best deal in the world from their point of view. Wouldn't it make more sense to be mad at every other employer for not proposing better pay, than the one that's offering the best pay for that worker?

Sure, but think about how humans drove others extinct. We never decided to "kill all wooly mammoths", we just wanted to use their meat / habitats for other things.

The correlation you mention seems noisy enough that I wouldn't want to bet my civilization on it.

Okay, well it seems funny to both claim that there is no good definition of agency, but also be sure that we've never demonstrated any hint of it in theory or practice. Planners, optimizers, and RL agents seem agentic to some degree to me, even if our current ones aren't very competent.

Yes, I agree that initially, AIs will be highly integrated, and their goals will probably at least appear to be somewhat aligned with those of human societies. Similar to human corporations and institutions. But human institutions and governments go off the rails regularly, and are only corrected because humans can sometimes go on strike or organize to stop them. I fear we will lose those abilities.

As a concrete example, North Korea was forced to allow some market activity after the famines in the 90s. If the regime didn't actually require humans to run it, they might easily have maintained their adherence to anti-market principles and let most of the people starve.