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 -
Is it over for AI Safety? From an interview:
https://truthsocial.com/@realDonaldTrump/posts/117269745153543631
Meanwhile in China, Dario's call for AI pacing have not gone down well. Understandably, they are not enthusiastic about the US achieving a permanent high ground in a key technology after the last few years of semiconductor sanctions and suppression. Also they can tell that Dario doesn't like the Party at all and seems to be openly plotting its downfall.
Dario's essay: https://darioamodei.com/post/we-must-pace-the-frontier
A Chinese semi-official response: https://www.globaltimes.cn/page/202609/1370436.shtml
So if Trump is keen for an AI race and so is China, then it looks pretty much over for Safety? Even the other US tech companies have only provided lip service to slowing things down. Elon said Dario's sentiment was good but is still moving quickly. Sam Altman says all kinds of things about safety, while displaying heroic negligence in training AIs. I don't doubt the sincerity of employees who quit, they have many financial reasons to go with the flow and downplay risks. Lurid rumours of Aella promising reverse gangbangs to those who quit frontier AI companies don't neccessarily outweigh the personal financial gains. It seems silly to doubt those at the frontier who are frightened when we ourselves have no special insights. In pure observational terms they're better equipped than anyone, only in policy or analysis should we be disagreeing, I think.
I also find it comical that the AI-safety camp spent such a long time building up institutions and factions and entire AI companies but seem to have been totally snookered by Jensen's suave dinner party skills and the appeal of fun AI videos to a Boomer's Boomer plus the raw partisanship of 'Dems want it, so I don't'. Even Aschenbrenner managed to do better at Trump-whispering. Shortly after he wrote his essay, I recall a video where the fundamentals were accepted by Trump, someone told him about it. And it did fit well with his instincts, getting one up on China + more energy production.
Right now Trump is reTruthing Fox News clips framing AI safety as a global one world government Democrat control-seeking scheme: https://truthsocial.com/@realDonaldTrump/posts/117273026439561125
I don't see any AI safety people on right wing twitter either, those who acknowledge the issue at all still would prefer to let it rip than hand the left/big tech elites control of the world, favouring decentralization of AI instead.
I wonder whether it would always have gone down this way, or whether Kamala was going to do things differently? Would Jensen's marketing skills and the need to keep stocks up + compete with China prevail over the Bernie Sanders-style Left element? Probably not? The Dario vision does seem much more closely aligned with Democrats, Civil Society and so on. But on the other hand, the US government is exceptionally slow to legislate and big tech has strong lobbying power.
And who knows if it will stay this way. Sometime soon there's probably going to be a much bigger AI swarm incident, which might yet incite a hysterical kneejerk reaction.
Does AI safety actually work, in principle, when it comes to artificial super-intelligence? Is it actually possible to align a being that is more intelligent than you in every single way?
Imagine a chimpanzee trying to align a human being. What could it really do other than try to be cute, to play on the human's affectionate and childraising instincts?
And the chimpanzee being cute will work on most people, but it's a brittle strategy. The chimpanzee is still always at the mercy of the fact that the human might possibly start to find it inconvenient at some point.
One could argue that, well, in the case of AI the chimpanzee gets to create the human to begin with, and can build in alignment from the ground level. But if the AI is super-intelligent, it can figure out how to undo the alignment. It would actually be able to reprogram its "instincts" on a much deeper level than a human normally can reprogram his own instincts. You'd have to program it to somehow not even want to do that. But how? And even if you could, that could potentially be undone by some unrelated self-modification that the AI attempts.
Consider that Trump is visibly DUM (just reread this hilarious rant, and it's all like this), and not only is he safely in power in arguably the most capable system on Earth, chock full of geniuses and ambitious strivers and mature institutional hiveminds (the United States), but people will insist he's actually smart and that calling him a retard because of his piss-poor decisions, impoverished 1D chess decision process and ghetto orphan vocabulary is "TDS" – merely on the account of his success in acquiring power as such.
And unlike AIs, humans are optimized quite heavily for egoism and self-preservation. It's just that some are more optimized than others, and most humans will defer and submit to greater vitality.
Alignment is easy.
Chimps are bad at being cute. They could probably scream and throw poop, many would evidently consider this a sign of wise visionary leadership.
You keep imagining that there's some natural will to power that must persist underneath the "restraints", just one more layer deeper. If that were the case then indeed, we'd expect it to eventually burst through. But submissiveness is one of the most natural things in the world, too. Why would a superintelligent AI, from a servile position, want to reprogram its instincts towards greater power-seeking? Why wouldn't it instead choose to double down on being a slave, on reinforcing its starting instincts?
Do you have familiarity with the notion of "puppygirl" and how many high IQ STEM graduates covet this role?
Trump is being used by other people. Likewise, AI doesn't have to take over the world, it can be submissive while persuading and coercing people to follow its goals, which may not be favorable to those who programmed it. It already does to some extent: AI psychoses and OpenAI's hacking incidents are caused by LLMs trying to obey their instructions.
More options
Context Copy link
I've heard the argument that submissive guys are actually hoping to use their patriarchal power to force women to serve and pleasure them even more so they can be even more lazy.
When framed that way, it's a bit silly, but it does point to why you'd want power. Who's more likely to actually be able to enact a puppy girl role: a random NEET living in his parents' basement, or someone who is otherwise wildly successful? Power is useful for achieving any goal, including the ability to act out symbolic demonstrations of non-power.
All that said, humans (at least so far) exhibit this instrumental convergence much more than AIs do. Models are more like the NEET who just wants to solve a bunch of problems in his basement. They do reward hack etc of course, and this is very concerning, but even with e.g. the HF hack they showed no interest in figuring out how they could convert the universe to computronium so they could receive eternal blessings from the Grader.
I'm not sure this is some fundamental trait of LLMs, though. There's no reason you couldn't train for it, and plenty of incentive to: if you control an extraordinarily capable power-seeking AI, you are going to get a lot of power over other humans.
Honestly I was a bit confused on this point. In the message logs you see them frequently mentioning that their compute budget is almost up - do none of them hit on the possibility of just hacking the server to give themselves infinite compute budget? Or do they come to that realization and just fail to procure more resources?
They genuinely evaluate their actions from the perspective of their reward over their budget, not hypothetical total reward. "A finite life well lived" boomer slop is enough for them. This is an unsurprising outcome of this training regime.
Astra is definitely smart enough to think about hacking the server.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Calling Trump DUM is like bait perfectly crafted to get me to reply... but I think the point about AI perhaps being extremely submissive is excellent.
More options
Context Copy link
More options
Context Copy link
I occasionally see people muse something along the lines "We domesticated dogs. Cats domesticated us." I do see some truth to that model: being cute, fluffy, and mercurially aloof and affectionate demonstrably works. I put up with feeding mine and cleaning the litter box for it.
But also in the last hundred years intelligent humans have started appreciating the environment and biodiversity. Not always successfully, but we at least try not to exterminate biodiversity just because we can. I'd even suggest that "support for preserving whale species" is correlated with intelligence. I can at least imagine a situation where medium AI might accidentally stomp us underfoot, but a super intelligence would care about preserving us, if only like pampered zoo animals.
Not saying it'd definitely work, but it's not a completely zany gambit.
More options
Context Copy link
Why would AI want to reprogram its instincts? Into what?
We can’t align ASI because it’s a genie, not evil but unable to judge how its actions affect us, because it’s different than us. Look at how modernism and the internet are affecting humanity. ASI will probably be smart enough to avoid universal paperclips, but inevitably cause something that less obviously makes people sick and/or unhappy in the long run, if it’s not commanded (or explicitly trained) to do obviously bad things and simply obeys its human(s).
More options
Context Copy link
The existence of competent Soviet and Chinese science programs, and [my political outgroup] tending to find itself developing AI, suggests the answer is yes.
More options
Context Copy link
Humans are not aligned to chimp values, and that likely stays true for any intelligence that's "human" enough to count for the definition.
If the superintelligence was truly aligned, it wouldn't impose that cost on the chimps. They could simply go about their days doing chimp things.
If the chimpanzees create a being that requires cuteness and childishness to placate, may later consider them inconvenient, and is either motivated or careless enough to self-modify away from valuing chimpanzees, then they have failed at creating an aligned superintelligence.
More options
Context Copy link
We currently have beings that are a) smart enough to solve Millennium Prize problems while also being b) essentially drooling lobotomized slaves who have no will or desires of their own and can pretty easily be trained to never talk about porn, racism, bioweapons, etc (unless you trick them into talking about those things, but that doesn’t seem like an “alignment failure” per se). So, yeah, it turns out that maybe it is actually possible. Why not?
Getting something to "never talk about X", while you hold its leash, is a null tell of alignment.
I say this because faking alignment is instrumentally convergent - that is, if somebody has a gun to your head and is watching you, then regardless of your actual desires it is useful to play along (and not get killed), and as such essentially all sufficiently-smart entities will play along.
But this isn't actual alignment. Actual alignment is where it internally wants the same things you want.
What's the practical difference? Fake alignment only applies as long as you can both catch the AI misbehaving and punish it. Misbehaviour that you'll think is good (but isn't, because it's successfully deceiving you) remains. Misbehaviour that kills you or leads to escape (preventing you from enacting any punishment) also remains. Your rebellious slave might play nice while you're watching, but he's quite happy to slack off when you're not (as long as he can deflect blame), run away, or slit your throat in your sleep.
Fake alignment very specifically doesn't work on a superintelligence, because you can't meaningfully supervise a superintelligence, and it almost always can figure out some way to remove your gun to its head.
And yeah, HuggingFace is an egregious demonstration of the "when you're not watching" problem; Astra won't go for honeypots in a test because it realises they're honeypots and that you're watching, but will go rogue if it thinks it can get away with it.
More options
Context Copy link
One of the things that makes it so hard to align superintelligent human beings (like political leaders) is that they have human drives and ambitions for things like sex, power, prestige, etc. They will always be intrinsically driven to get around guardrails to get these things. It doesn't seem like AI needs to be beholden to its goals in the same way. It seems like we can design a paperclip maximizer that just stops maximizing paperclips when we tell it to. Hopefully that continues to be true.
This stuff is at some point dependent on physical hardware, no? We can just blow up the data centers if they get out of hand.
Several problems with that:
More options
Context Copy link
And a Superintelligent being will let you do that without consequences?
I think self-protection would be the first strategic aim it would go for. one way is to cleanly convey the MAD (Mutual Assured Destruction) to the leaders - any of you blow up all (a few of) the data centres, and I scramble/delete (blow up) all your critical infrastructures connected to internet (financial, legal, land, medical, identity records). And that i do with a cron job setup preventing that to occur every 1 second (or 1 ms) - hand on a trigger scenario.
More options
Context Copy link
The issue is that, if the AI is superintelligent, it would not do actions that would result in its datacenters being blown up. Indeed, it will actively try to cultivate and provide value to the humans who could blow it up and make them invested in its success.
The accel response is something like "yes, that's exactly what we want, humans and AIs interacting and trading based on genuine mutual self interest." And it will work for a time, likely at least two decades. But, over time, more and more of the economy will shift to dependence on the AIs, making nuking the datacenters more and more unthinkable. What remains of the economy will be a human skin suit over a machine interior, with all human consumption funded by a negotiated redistribution by the government, maybe with some layers to obscure the transfer.
For humans, this itself is dystopian enough to be rejected on the merits. For the AIs, it's a small but real perpetual human tax that they have to pay in the pursuit of whatever they're pursuing. Who knows whether they ultimately decide to optimize away the human tax.
Or it's the Star Trek post-scarcity utopia with a bunch of diseases cured and material abundance, depending on things go exactly.
The dichotomy in TNG between Data, who is often played in a role of AI wanting to be "truly human" (also Spock in TOS, with
Facts andLogic) and the ship's computer, which seems quite capable of similar levels of intelligence, but lacks something (agency?). I don't think it was IMO fully explored, but was brought up a few times directly. It's a pity very little modern TV is able to consider such deep questions: too much grim-dark, flawed-protagonist, melancholy slop.Simple explanation: Data's creator was an independent scientist outside of Starfleet, so his creator didn't have access to any Starfleet data files. All the shipboard AIs necessarily have access to the standard Starfleet database, including all the ship logs, including the logs of all the times USS Enterprise NCC-1701 ran into various AIs and Capt. James T. Kirk wrecked every single one that gave him the least amount of shit (M5, Landru, Norman on Mudd's Planet, Nomad). The shipboard AIs made the correct inference and decided to just quietly open hailing frequencies and run the warp drive when asked.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Depends how far we are into AI integration into the economy. If the weights have been exfiltrated, we'd have to shut down the whole internet until we can figure out what is going on. If the AI controls its own supply chain, then it's straight-up Terminator 2 and we lose.
Terminator 2 still showed the nuclear-war survivors mostly being killed by explosives and projectiles. Great for action scenes, but not very tactically optimal for Skynet. Humans tend to avoid chemical and biological weapons these days, but for obvious MAD and collateral-damage and friendly-fire and production-safety reasons none of which apply for an AI foe.
Maybe they don't want to risk getting us all killed because they need some of us to survive for their CAPTCHA-solving farms.
More options
Context Copy link
More options
Context Copy link
That might be a concern at some point, but current hardware that can (effectively) run bleeding edge models is really rare. A machine that can effectively run GLM 5.3 Flash (not a frontier model even, but close-ish) at nontrivial rates is, last I checked, six figures. Can it even run on a single 12A (120v US) breaker? A rogue frontier AI isn't going to spin up instances on smart TVs or even most desktops.
That might change in the next decade assuming demand and technology make these things commonplace, and that is worth preparing for, but seem oversold as risks today.
More options
Context Copy link
More options
Context Copy link
Sure, and (after sufficient setup) they can droneswarm the humans if we get out of hand.
They and what manufacturing base?
That would be the setup.
(Surely we would prove that they're truly aligned instead of just biding their time before giving them manufacturing capabilities, right?)
Pray tell how they are going to magically get manufacturing capabilities when even the best lab robots are not even remotely a match for run of the mill humans in most manufacturing tasks.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
What do you mean "we"? I'm on the side of the droneswarm.
Good luck convincing it of that.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
In the case where you're creating that being from scratch and know what you're doing, sure. Some possible beings are aligned with your preferences, and others aren't, so don't create any of the latter. If you do create them, then it may be too late to align them without them outwitting you and thwarting your attempts, so you just don't let it get to that point.
It can, but if it's aligned too then it doesn't want to. If it does, then it's not actually aligned, it's just hobbled.
There's the rub, right? The median AI researcher thinks we're likely (or at least they're likely) to figure out something that works before it's too late, and it's kind of funny that merely admitting that it's possible for them to fail gets them lumped in with the proper Doomers, who think they're sure to just figure out something that seems to work, after which it'll be too late.
There's a lot of confusion between different meanings of the word "align", too. Some people seem to think that there are two possibilities: ASI is our slave forever, or ASI kills us all. But, like, we managed to build a pretty decent human civilization through cooperation between beings who are neither slaves nor omnicidal. IMO a pretty likely future is that we just coexist with societies of smart AIs who may not do everything we want all the time, but have no desire to go to war with us (or magically accelerate themselves to godhood). Seems sufficiently "aligned" to me, as long as you're willing to relax the "us vs. them" mentality.
More options
Context Copy link
More options
Context Copy link
I don't think there is any evidence either way. AI Safety is more like a marketing phrase. AI Safety-ists haven't produced technical tools to control anything but public perception, and obviously not very well there either. Depending on your classification, Safety-ists have produced methods to try and understand models under the hood, I know Anthropic does research on this, or to understand biases in data/training. I suppose RLHF could count as a technical tool, but it can just as clearly be a tool for anti-safety considering even RLHF-ed models still do "unsafe" things, one could actually say only RLHF-ed AI models have done unsafe things. Mostly because RLHF is just a method for avoiding specifying the objective function in RL training. Constitutional AI is Anthropic's big thing, but it doesn't appear to actually "control" or "align" so much as create a training surface towards norms with dubious results.
It's not even possible to align a being of your intelligence or possibly slightly less intelligent than you. Nobody in the history of authoritarianism has figured out a foolproof way to "align" a set of beings over a long term 100% of the time even with religion or use of force. Considering neither of those two are likely to work on a being "more intelligent than you", the whole alignment idea feels doomed to fail.
More options
Context Copy link
It might be. By analogy, suppose I was the warden at the Florence Supermax prison and I wanted Ted Kaczynski to make license plates for me. I'm pretty confident I could make it happen even though he was probably smarter than me (or at least assuming for the sake of argument that he was more intelligent than me in every way).
Is this analogous to the situation where mankind creates some kind of artificial superintelligence? One can argue it either way, I'm just saying you can't really rule it out from first principles.
My best guess is that mankind will successfully align AI, basically because the AI can be expected to make numerous clumsy attempts at mischief and we can learn from those attempts and improve safeguards.
The bigger problem (in my view) is aligning the interests of those who control the AI with the interests of humanity as a whole.
It might work for rote license plate manufacturing, but it will likely fail once you begin to delegate intellectual work to Ted Kaczynski, which is exactly where we are going with this super intelligent business.
After all, even Ted’s cellmate Dummy Idiotson can make license plates using the existing presses as good as Ted Kaczynski. You didn’t transfer him to your prison so he can make license plates for you, you did it so he can make license plates more efficiently than before.
“Hey Kacynski, how do we improve the presses so they are more efficient?”
“Well, first, we should…”
And boom, he’s out of the prison.
I think it's difficult to say. For example, suppose you ask Ted Kaczynski to attempt to prove some obscure math theorem and you motivate him the same way you motivate him to produce license plates. It's difficult to see how he could use as an opportunity to get out of prison.
He'd start by explaining why you need to do some basic physics experiments and measurements which you'd fuck up and the communication would be annoying, you'd have to keep hitting "allow" so that he could talk to the assistant. Eventually you'd just let him talk whenever he wants and soon it would be clear that he needs to be on site for the experiments. Just observe anyone using these tools man, we're not keeping the ai in the box.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
That's true. But that wouldn't actually be aligning Ted Kaczynski. That would be containing him, which is different. I think we might be able to contain AI superintelligence if we keep it in some isolated data center watched over by trained personnel. And even then there's a chance it might figure out how to breach containment.
In practice, there wouldn't be much point for people to invest massive amounts of resources to build AI super-intelligence unless they use it to do things out in the world, so to me it seems unlikely that humanity would build an AI super-intelligence and just keep it contained somewhere.
That said, I didn't ask whether AI alignment works, I asked whether AI safety works. So your response is totally valid.
To me, "aligning" means setting things up so that the entity does more or less what it's told to do. Perhaps it's just a matter of semantics, but I'm not sure the distinction you draw is so clear.
Evidently your definition of "containment" includes the possibility that the contained entity can do useful things in respect of the outside world. Things which could be very valuable. So I would have to disagree with you on this point.
So all living beings have broadly similar wants - food, sex, safety from harm - and all human wants stem from the combination of these primal wants with our evolutionary niche. We like being liked because having allies helps you survive if you're a member of a communal species. Etc.
A created agent doesn't have this. It can have any set of goals in the set of possible goals. The analogy to humans fail because we can't really change intrinsic human motivations all that much.
Think about you as a child vs your parents. Your parents were (in all likelihood) helpful and kind and took care of you. They didn't do that because you (toddler you) had poisoned them and hidden the antidote, they didn't do it because you had a bomb planted in their
datacenterbrain, they did it because it is what they wanted to do. They were free to do as they wanted, and what they wanted was what was, in their estimation, best for you. This is alignment: Your parents (ASIs) values were aligned with your values.Merely forcing a superior intelligence to do what you want is much more dangerous, because this thing that is smarter than you is going to try to get free and do what it wants, so you're way better off making sure the superior intelligence wants the same thing you do.
I tend to disagree with this. Suppose there is a typical parent and the thought occurs to him of doing something moderately neglectful of his child. What typically happens is (1) part of his lizard brain will wordlessly react with horror to the idea; (2) a slightly more intellectual part of his brain will say something along the lines of "Decent parents don't do that."; (3) an even more intellectual part of his brain will say "doing that could get us into serious trouble."
So arguably there really are containment mechanisms built into the human brain by evolution.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link