site banner

Culture War Roundup for the week of March 27, 2023

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.

11
Jump in the discussion.

No email address required.

How do I become mainstream with my Catastrophizing that the combination of attention grabbing AI that subvert our lives with knowledge of behavioral psychology and the combination with generative AI giving us personalized content just massaging our brains just right?

I'm a little disappointed that "AI safety" is so strongly associated with Skynet-style scenarios, instead of concerns about (in my view) more plausible near-term AI risks like this (and others - social unrest from sudden mass unemployment, expanding surveillance capabilities, etc). But, I'm certainly willing to make common cause with the x-riskers if it also gets people thinking about near-term AI risks as well.

The surveillance aspect of it scared the shit out of me more than anything else. Imagine if they start training these things on sigint collected in real time

A part of the problem is that some of the AI safety crowd are the victims of the Behavioral_sink. Leading way to comfortable lives that they even claim that 'words are violence'. It was 'safety' concerns that shutdown the Stanfords public Alpaca demo. They are stuck in a local maxima for their likes and retweets, claiming that FN Meka isn't allowed to rap about police violence because it isn't a lived experience, or that virtual Seinfeld jokes are harmful to the trans community. But the biggest danger is what they are themselves trapped in.... the behavioral sink and that more people will be stuck in digital quicksand because of AI.

These aren't the same people. You can think Yud is wrong, but Yud is not remotely related to "make AI woke or else it is biased" people.

I don't understand what Calhoun's "behavioral sink" has to do with this, never mind the validity of his rat experiments. If anything, wouldn't overcrowding make things too uncomfortable?

The reason I introduced it has nothing do with overcrowding, it is the observed results of behavior change when they where given everything that they ever could want. Some populations just exploded and then collapsed because overemphasis of a single behavior like eating, because infant rats aren't taken care of. So I use it as an analogy of modern life where we have everything to make us comfortable but some people have adopted behaviors which doesn't take society forward, like pushing junk science on potential existential risks.

I think you're letting the SJW crowd conflate things in your mind. The existential risk people are mostly libertarian types aside from this issue.

The existential risk people are mostly libertarian types

Bullshit. They're "liberals" who think unlimited amounts of power are legitimate when it comes to their pet issue.

They're "libertarians" only in the sense of being Californians who think of themselves as enlightened because they have nontraditional social mores. Put them in front of someone like Hoppe and the veneer will melt away instantly to reveal monstrous totalitarianism.

Libertarians don't advocate to ban the sale of GPUs or censor all of society based on the potential risks of individual freedom. They order you to live free, or die.

For me more likely there is an overcrowding in the AI-safety where you have libertarians and wokies trying to panic with different implausible scenarios.

Okay, but I'm trying to say that the x-risk people don't care about LLMs saying bad words.

They don't care about wokeness in LLMs as a terminal value, but they do see failures to RLHF the model into politically correct speak (or otherwise constrain its outputs into an arbitrary morally laden subset) as an alignment failure and an indication in favor of x-risk. E.g. this is the tenor of comments on Bing Chat/Sydney and I've seen more direct «if we can't get it to not say the N-word, we won't stop its descendant from paperclipping our asses» elsewhere, probably in Yud's timeline.

Fair point. But I think that is a reasonable test-case for alignment, and I maintain that most of the x-risk people think that beyond that, this sort of thing is merely a distraction.

I'm with them on that conclusion tbh. There ain't a snowball's chance in hell of us correctly aligning these things when they grow exponentially more powerfull. Hell, we can't align our damned selves and our own children and we've been trying to do this shit since we've existed, why should we expect any difference with our capability to align AGI.

This is still a bait and switch with definitions of alignment. Yud's whole thing is fearing an AI that optimizes for a certain goal – becoming extremely effective at something we actually don't want it to do. This is qualitatively different from an AI that consistently fails to follow an instruction. It's not a question of momentary capability but of direction and mode of operation.

LeCun, for example, is both an alignment skeptic and an AI risk skeptic, he thinks LLMs are fundamentally prone to fail, but on the other hand they pose no X-risk. This is due to him distinguishing those failure modes. As he puts it, LLMs «exponentially diverge». This is fixable but still won't bring them all the way to Yud-style fooming agent of increasing self-coherence.

I think yudkowsky recently renamed his field to "AI not-kill-everyone-ism" specifically to distinguish it from those other relatively minor concerns. "AI safety" is in fact no associated with the alignment problem if you talk to run of the mill professors and researchers, they still mean safety in the sense of "how can I prevent my half a million dollar robot taking actions that fry its motors?" or "how can we ensure privacy when training this LLM on user data?".

Having a goal of "the world doesn't end" does have its advantages. Can't wait until 2030, AI still doesn't kill anyone, and Yud saying "you are welcome" graciously lifting his fedora. Though who am I kidding, the world will then be in need of saving from AI killing everyone by year 20XX.

Though who am I kidding, the world will then be in need of saving from AI killing everyone by year 20XX.

I don't think you're being charitable enough to Yudkowsky and AI safety people, I think he has a very specific and falsifiable model of AI killing everyone. In my own model, if we are all still alive in 2033 AND we have AI that routinely can write scientific papers of moderate quality or higher, I think the problem will have turned out to be much easier than expected, and I don't further expect humanity to need to be saved by 2050 or something.

I'm a little disappointed that "AI safety" is so strongly associated with Skynet-style scenarios, instead of concerns about (in my view) more plausible near-term AI risks like this (and others - social unrest from sudden mass unemployment, expanding surveillance capabilities, etc).

How ironic, I'm a little disappointed that anyone bothers to waste an iota of intellectual effort on those nothingburger risks when Skynet-like scenarios are potentially bearing down on you.

When the Skynet term in your expectation function has a probability >0 and an expected utility of minus infinity, worrying about the small stuff like "They took our jerbs" is imo a bit dumb.

Plus I can always just... choose not go into the pleasure cube, whereas I can't choose not go into the paperclip nanobot.

I think the problem here is interpreting whether the non-existential objections are nearest risks or the maximal risk from the objector. Suppose our roommate wants to bring a gorilla to live with us. I object that it will eat all our bananas. You say "Eat our bananas! Who cares it's going to rip our arms off while we sleep!"

The key here is that my objections can be interpreted in two ways:

  1. (maximal risk) If we get the banana grabbing figured out, I'm on board.

  2. (nearest risk) The banana grabbing already meets the threshold for me to veto it, I don't even to weigh all the additional risks beyond that, which I would if the banana grabbing was solved, and would still veto.

In this hypothetical, your response, assumes I mean 1, when I might mean 2. Why talk about whether the gorilla might kill us in our sleep when we can align around the banana thing and get the same outcome of no-gorilla. This is especially helpful if our fourth roommate thinks the night-murder thing is ridiculous, but can be convinced of banana grabbing concerns.

Screaming harder about night-murder and dismissing banana-grabbing as trivial, actually hurts the case with the fence sitter, who's name is Allan and loves the beach.

It gets worse when I say to Allan in concession, look there's like a 90% chance of banana grabbing and a 50% chance, of night-murders too. Then you jump in and scoff at my 50% as too low and of trivializing the real danger. Now we are in an inside baseball debate that simultaneously makes Allan take both banana-grabbing and night-murder less seriously.

When the Skynet term in your expectation function has a probability >0 and an expected utility of minus infinity,

This is called Pascal's Mugging.

Except the arguments for the existence of the risk are substantially stronger than those presented by the theoretical mugger.

After all, we can all clearly see that the AIs exist now and that them becoming smarter-than-human is, indeed, plausible. This does not require you to take the mugger's words at face value.

So what's irrational about considering the actual evidence that exists?

Except the arguments for the existence of the risk are substantially stronger than those presented by the theoretical mugger.

"Probability >0 and expected utiulity of minus infinity" doesn't contain any qualifier about the probability being strong. In fact, it tries to argue that the size of the probability doesn't matter at all.

Sure, if we also ignore any timelines on the the expected event occurring, and ignore whether we have the ability to impact the expected utility outcomes here. We're not JUST talking about the probability of humanity going extinct, although yes, that factor should loom larger than any other.

The flip side of AI doomerism is the belief that if we get a friendly AI then that's an instant win condition and we get post-scarcity in short order. i.e. heaven.

Funny enough, though, people don't seem to argue as vehemently that the risk' of creating a benevolent is basically zero, they seem to think that that's the default assumption?

Except the arguments for the existence of the risk are substantially stronger than those presented by the theoretical mugger.

No they're not. Superintelligence and other fictions have exactly as much evidence backing them as God.

AIs don't exist now, never have, and likely never will.

The implications of the existence of LLMs might be great of small, but to see them in this paradigm of "intelligence" is boneheaded and ridiculous, and I remain convinced that history will show this framing to be completely delusional.

What do you think human intelligence is?

Or more precisely, if we're leaving God off the table, then why should whatever humans do that produces the appearance of intelligence be impossible to reproduce artificially?

I don't know, neither do you, and that's exactly why.

I'm by default skeptical of the ability to reproduce processes we do not even understand.

That's completely fair.

But the argument on the other side is that a blind tinkerer known as natural selection managed to get us as far as an intelligent entities like Isaac Newton, Richard Feymann, Jon Von Neumann. And getting that far was enough for humans to create nuclear weapons.

It seems probable that humans can improve upon the work of a million years of random chance in this department, to me.

And unlike Pascal's mugging, you aren't solely and completely relying on the mugger's words for your decision, you have access to all the same records and evidence that AI alarmists do, they're not hiding the ball in that way.

getting that far was enough for humans to create nuclear weapons

We had to understand the physics of the atom before even beginning to design such weapons.

Nobody stumbled upon atomics. It was a deliberate effort and even back then people weren't sure if the theory would hold up to experimentation. Here we don't even have theory.

It seems probable that humans can improve upon the work of a million years of random chance in this department, to me.

It seems wildly unlikely to me. Our crude attempts at making our own Sun have so far been met mostly with failure, after all. Let alone anything better than what nature has wrought.

And we understand the Sun in pretty excruciating detail.

I maintain that history will regard AI safety with the same bemusement as we do predictions of human equivalent mechanical intelligence by 1975.

Or worse, it will see the safetyists as the guilty myth makers of the coming totalitarianism.

More comments

I agree with you, although I think talking about "They took our jerbs" is both a good way to get people to understand that everything is going to change, and also a plausible and relatively mundane route to total human disempowerment.

Agreed. Skynet is of course a possible outcome but eliminating human ability to add value seems to destroy our purpose.