This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.
Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.
We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:
-
Shaming.
-
Attempting to 'build consensus' or enforce ideological conformity.
-
Making sweeping generalizations to vilify a group you dislike.
-
Recruiting for a cause.
-
Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.
In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:
-
Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.
-
Be as precise and charitable as you can. Don't paraphrase unflatteringly.
-
Don't imply that someone said something they did not say, even if you think it follows from what they said.
-
Write like everyone is reading and you want them to be included in the discussion.
On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.
Jump in the discussion.
No email address required.
Notes -
There's one way I could maybe see us having problems recreating some facet of modern tech. That is, indeed, a nuclear war, and the resulting radiation causing the most advanced computers to crash often (since modern RAM/registers operate on such exact precision that they can be bit-flipped by a single decay). Even then, though, there are ways and means of getting around that; they're just expensive.
Ord indeed takes an axe to the general version of this argument. Main points: 1) in many cases, resources are actually more accessible (e.g. open-cut mines, which will still be there even if you ignore them for 50 years, or a ruined city made substantially out of metal being a much easier source of metal than mankind's had since native copper was exhausted back in the Stone Age), 2) redeveloping technology is much easier than developing it for the first time, since you don't need the 1.0, least efficient version of the tech to be useful (e.g. the Newcomen atmospheric engine is hilariously inferior to what we could make with even similar-precision equipment). There are a whole pile of doomsday preppers who keep this sort of information in hardcopy in bunkers; we're not going to lose it. And, well, 1700s humanity (knocking us back further than that even temporarily would be extremely hard, because pre-industrial equipment is buildable by artisans) is still near-immune to natural X-risks; I'm less convinced that 1700s humanity would survive another Chicxulub than I am of modern humanity doing so, but that is the sort of thing it would take, and shocks that large are nailed down with low uncertainty at about 1/100,000,000 years.
If you really want to create a scenario where being knocked back a bit is a problem, I think the most plausible is something along the lines of "we release some horrible X-risk thing, then we go Mad Max, and that stops us from counteracting the X-risk thing". Global warming is not going to do that - sea levels will keep rising, of course, and the areas in which crops can be grown will change a little bit more, but none of that is too fast for civilisations to survive. (It's not like you're talking about 1692 Port Royal sinking into the ocean in a few minutes; you're talking about decades.) Most of the anthropogenic risks are pretty fast, so they're ruled out; we're dead or we're not. Life 2.0 is about the only one where I'd say "yeah, that's plausible"; that can have a long lead time.
Dysgenics is real but not very fast, and it's only plausibly been operating for what, a century, and in only about half the world? This isn't going to be the end of the world. Flynn effect would be wiped out in apocalypse scenarios, of course, but we haven't eroded the baseline that much.
And to zoom out and talk about X-risk in fully-general terms, I'll say this: there are ways to mitigate it that don't involve opening the Pandora's Box of neural-net AGI. Off-world colonies don't need AI, and self-sustaining ones take an absolute sledgehammer to every X-risk except AI and dystopia (and aliens and God, but they're hardly immediate concerns). Dumb incentives for bio research can be fixed (and physics research, if and when we get to that). Dysgenics yields to PGT-P and sperm donors (although eugenics has some issues of its own). Hell, even GOFAI research or uploads aren't likely to take much over a century, and would be a hell of a lot safer than playing with neural nets (safer is not the same thing as safe, but fine, I agree, keeping AI suppressed on extremely-long timescales has issues). "We must do something" does not imply "we must do this".
All of Ord's calculations seem to assume highly efficient coordinated conspiracy running the world instead of the slowly decaying shitshow one should expect from post-collapse civilization, but I guess that's par for the course.
Why do you assume GOFAI or uploads are safer? From what I can tell, this sort of reasoning is only grounded in aesthetic distaste for statistics, on in mushy intuitions like «much weight, very black box, there be shoggoth» like that fraud Leahy implies. Well, we can sparsify models and make them structurally simpler, it seems, but what of it? The causal structure of a mind worth the name cannot fit into human consciousness, and it's improbable that there are theorems that give sufficiently secure bounds for its qualitative behavior; if there are, they could as well hold for NNs, because why the hell not, they are differentiable functions. GOFAI either doesn't work at all (there are good reasons to expect it to never work, or to require neural nets to assemble) or is more risky in terms of having sharp transitions and giving us unbounded optimizers. Indeed, all of Yud's fears were invented basically for GOFAI or are paradigm-agnostic; he never expected NNs to work, not without intractable compute expenditure on the scale of simulating historical evolution, and so the question of their – now supposedly inevitable – misalignment was moot. Likewise I don't see what makes uploads inherently safe but doesn't hold for NNs. Are they too fat to be a threat or what? They'd still be data, still amenable to self-editing, still capable of accelerated and inscrutable interactions, eventual economic superiority over baseline humans, inhumanity. See Pantheon, it's as good a doomsday narrative as any.
Reminder: having received millions of dollars in donations, Yud has turned his AI safety organization MIRI into a sinecure for mathletes preoccupied with weird self-referentiality studies and esoteric decision theories. What little alignment-related investigations they had were nonsensical to the point of casting serious doubt on coherence of their conceptual apparatus:
Insane asylum stuff. Unsurprisingly, Christiano, ever the gentleman, describes the situation such:
No, really, what do you have against neural networks? I think it's tremendous luck in terms of safety that we got them to work and can approximate/interpolate/extrapolate desirable behavioral properties directly, instead of languishing for a few more centuries and praying that our Enlightened Infra-Bayesian Lords didn't make an oopsie with the sign somewhere and covered it up like another Leverage, or the Wuhan affair.
The view I'm coming at this from is: humans have a moral skeleton, innate hardwiring that allows us to learn morality and believe it (as opposed to mimic it). This is highly instrumentally non-convergent and probably needs to be coded into an AI directly; gradient descent on output will only produce lying psychopaths mimicking morality.
GOFAI has some hope because we could code morality directly. Uploads have some hope because you're uploading the hardwiring (whether or not you understand it). As I said, this does not equal safe, in either case; as you say, GOFAI has a lot of potential pitfalls, and uploaded people would be so far out of evolutionary environment that their remaining sane is far from assured.
But I'm not seeing any hope of success on non-uploads without the ability to look inside the box. This is because "is moral" and "is pretending to be moral successfully" have identical output except in situations where dropping the pretence is worth it i.e. situations where there's a high chance of you losing control upon betrayal. Interpretability might pull a rabbit out of the hat (I put it at about 3%, which is better odds than Yudkowsky gives), but I'm not very confident; to me, P?=NP notwithstanding, it seems like the difficulty of determining whether spaghetti-code does X is generally at least as high as the difficulty of writing code that does X, which implies that making safe NNs is at least as hard as writing GOFAI.
I suppose we have, to some extent, but it can't be all that robust. Tons of species are psychopathic by our standard, and of course this standard exists to distinguish humans who don't fit it. So it's more like a slight inductive bias, in the same way we have biases to learn to navigate in 3D space and prefer sugary foods. Biases of the algorithm can be substituted with biases in the data.
I don't see why that would be true. Indeed, I do not see why gradient descent wouldn't be much better of learning deep cognitive regularities including morality. You seem to hold that morality is something essential, some set of terminal value-influences, but why is that true for morality and not any other aspect of our cognition, both instrumentally worthwhile and «instrumentally non-convergent» ones? Every part of our decision-making feels profoundly qualitatively colored for us.
Why is "coded directly" better than learned? The major reason we're doing this stuff is that it's vastly better at generalization,
Sorry, this looks like a vibe-based argument, where neural nets deceptively "imitate" and hard code is "good and honest". It's all algorithms. Inasmuch as human minds are computable, our morality is an algorithm too.
What good would that do? It'd break OOD just the same, and if it didn't break, it'd be rewritten or worked around by the purported daemon of optimization.
Reminder that LLMs cannot learn to do the «treacherous turn» because the dreaded SGD mercilessly ablates cognitive routines that do not contribute to decreasing loss in training. This, of course, holds in the general case.
But even beyond that, outputs may be similar but activations aren't, we know how to look at activations, and we know there are differences between the model subjectively evaluating its output as true or false.
No, generation is always vastly simpler than classification unless you require classification that reconstructs the process of generation, of course.
I've long held that Yuddist program is, in addition to all else, an attractor for a particular variation of anxiety/OCD disorder: fetishization of «being in control», of making thoughts play by «proper rules». But it's fetishization because it doesn't really work, it pursues ghosts, precisely the deceptive external form of reliability. You gain clarity of ideas by testing them against evidence, not by being real suspicious of disconfirmations.
Pitts wrote that his depression might be “common to all people with an excessively logical education who work in applied mathematics: It is a kind of pessimism resulting from an inability to believe in what people call the Principle of Induction, or the principle of the Uniformity of Nature. Since one cannot prove, or even render probable a priori, that the sun should rise tomorrow, we cannot really believe it shall.”
This malady is to be treated, not nurtured.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Sure, but of course such measures being possible doesn't mean they'll actually be done.
This seems like too much certainty about the nature and difficulty of the task, which in turn influences whether significant delay actually increases the odds of success. For instance, if we turn out to live in a universe where superhuman AI safety isn't that hard, then the important thing is probably that it be done by a team that considers it a serious concern at all. Right now the leading AI company is run by people who are very concerned with AI alignment and who founded the company with that in mind, if we ban AI development and then the ban gets abandoned in 30 years there's a good chance that won't be the case again.
A candidate for such a universe would be if it's viable to make superintelligent Tool AIs. Like if GPT-10 can mechanistically output superhuman scientific papers but still doesn't have goals of its own. Such an AI would still be dangerous and you certainly couldn't release it to the general public, but you could carefully prompt it for papers suggesting more resilient AI alignment solutions. Some have argued Agent AIs would have advantages compared to Tool AIs, like Gwern arguing Tool AIs would be "less intelligent, efficient, and economically valuable". Lets say we live in a future where more advanced versions of GPT get routinely hooked up to other components like AgentGPT to carry out tasks, something which makes it significantly better at complicated tasks. OpenAI just developed GPT-10 which might be capable of superhuman scientific research. They can immediately hook it up to AgentGPT+ and make trillions of dollars while curing cancer, or they can spend 2 years tweaking it until it can perform superhuman scientific research without agentic components. It seems plausible that OpenAI would take the harder but safer route, but our 2050s AI company very well might not bother. Especially if the researchers, having successfully gotten rid of the ban, view AI alignment people the same way anti-nuclear-power environmentalists and anti-GMO activists are viewed by those respective fields.
Regarding talk of 100-year bans on AI while people steadily work on supposedly safer methods, I'm reminded of how 40 years ago overpopulation was a big mainstream concern among intellectuals. These ideas influenced government policy, most famously China's One Child policy. Today the fertility rate is substantially reduced (though mostly not by the anti-overpopulation activists), the population is predictably aging, and...the plan is completely abandoned, even though that was the entirely predictable result of dropping fertility. Nowadays if a country is concerned with ferility either way it'll want it to increase rather than decrease. Likewise the eugenics movement had ambitions of operating across many generations before being erased by the tides of history. In general, expecting your movement/ideas to retain power that long seems risky seems very risky.
More options
Context Copy link
More options
Context Copy link