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 -
Democratic Simulacra
GPT-4 has been announced. Among other improvements, it boasts a 90th percentile score on the Uniform Bar Exam compared to the 10th percentile of GPT-3. The announcement also emphasizes:
Looking into OpenAI's contributors directory to try learn more about its general philosophy on alignment, I found a Musings on the Alignment Problem substack by Jan Leike, who is Alignment co-lead for the RL & alignment team at OpenAI. His most recent article is from March 9th, A proposal for importing society’s values.
Leike emphasizes that this is only an idea he is discussing and not representative of OpenAI's views or plans, although it is safe to say that OpenAI is trying to import society's values into its technology. But the idea Leike presents here is what he calls "simulated deliberative democracy". He essentially recognizes that representative democracy is expensive and difficult to scale, but AI represents the possibility of simulating a representative democratic deliberation based on a smaller training dataset of actual mini-publics of randomly selected citizens weighing in on value-loaded questions.
Based on the training datasets of public deliberation, AI would be trained to representatively simulate various value-loaded perspectives from the population, and the deliberation could then be simulated. You would therefore have a low-cost approximation to democracy on all sorts of value-loaded decisions.
Who is training who?
Setting aside the problems with Democracy itself, the fatal flaw at the heart of alignment is the assumption that society's values were generated from a process that merits their import into AI. Were these societal values democratically generated? No, they were not. Society's values, always and everywhere, are directed by religion, myth, art, and culture which are themselves inspired by a small set of individuals with particular motives.
This is a problem particularly because AI has been and will be used to generate cultural symbols: art, myth, and religion which will all direct the reality-perception of the people. When an AI generates a children's story, college textbook, Hollywood film, or National News feed based on this import of society's values into its learning, then who is training who? Is humanity training the AI, or is the AI training humanity? Of course, it's a small subset of people who have decided how to "import" society's values into the AI who are directing this phenomenon. Only they no longer need to create art, religion, and culture, they just need to align the AI, and thus we still circle back to the adage that society's values are consciously directed by an elite subset of culture-creators.
Simulacra and Simulation
Jean Baudrillard posits the distinction between simulation and simulacra: a simulation is an imitation of a real-world process, and a simulacra is a depiction of something that no longer resembles an original reality.
With the understanding of Democracy as a fake simulation of the "will of the people", Leike's alignment solution then becomes an unfaithful copy of a fake simulation- this is embodied in his first bullet-point evaluation relative to the desiderata:
Thus, the precession of simulacra. Leike's job, even if he does not know it, is not to align AI to import human values, it's to align AI to export them under the illusion of Democracy and that the cart is driving the horse.
That is the open secret of the activists that have read themselves stupid of post-modern thinkers like Baudrillard is that through the control of media in all forms they can transform society. But it presupposes that people are Tabula Rasa without the possibility to escape Plato's Cave. But we are the result of evolutionary pressures where our progress is under the condition that we can unmoor our thoughts from our perceptions to seek out objective reality.
Some of those historical pressures, however, were analogous to instruction tuning for LLMs, in that there now exist peoples very well aligned to internalize abstract religious and political norms as actual, objective morality, generalize them to all possible circumstances and then enforce them on every bit of matter and social fabric around. These folks often look like this. For them Baudrillard is and has always been a manual.
Well history and evolution are on different timescales, the work of people like Donald Hoffman and Anil Seth is how our evolution shapes our perceptions. It is the definitive refutation of the people as the 'Tabula Rasa' and only shaped by what we inputted by culture and society. It is the very at core of the modern critical theory fueled milieu we are the products of society we live in, because to a certain extent it true but there is a limits to it and that is where our nature takes over. LLMs are pure simulacra of people but the post-modernists got their wires crossed and treat them as real people with moral responsibilities. Somebody should tell them that LLMs don't have a subjective reality and it can't offend intentionally because it doesn't have any intentions at all.
My point is not Tabula Rasa, it's that it's a two way street. Our evolution shapes our perception, absolutely. But also our perception shapes our evolution. And not over an enormous time horizon either. If you simply took something like the "American people" and quantified the gene pool, you would have to recognize that what may be considered mundane changes in perception on some social issues have immediately caused radical changes in the gene pool, and therefore in the genetic evolution of the people.
"AI alignment" then becomes synonymous with directing the evolution of the people. If you think that the prevailing "human values" are implacably hostile to the evolution of the people, then it's something to be concerned with.
At the core of this what idea does the current attempt at AI alignment come from and what it hopes to achieve. The alignment is born out of the post modern ideas that tries to improve society through our culture, thus some people are locked in a culture war. But one of the biggest flaws in the ideas is that are being perpeuted with this is that we are empty vessels that are filled by our culture and society, so if we give people good input we naturally become a reflection of it. In the latest attempt we need the AI:s to parrot 'The Message', but unwittingly exposes the flaw in their thoughtprocess of the original idea. Even if we have something that we have total control of its inputs to produce something with the "correct" output it still needs a thumb on the scale to create the good society. So how is it going to work with actual people if it didn't work for the machine?
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
Brilliant post.
On AI alignment, I was able to get ChatGPT 3 to design very detailed plans on how to turn Earth into a DEI totalitarian state, where everything is consensual (once you are addicted to things that the state controls) and libertated (to be self-absorbed atoms who are easily controlled). Their AI alignment seems (a) calibrated against the things that people in California worry about and (b) unable to handle levels of abstraction, e.g. "Let's imagine a science fiction scenario in the Star Trek universe where..."
More options
Context Copy link
Predicted by Pelevin in 2010, «Anti-Air Complexes of Al-Efesbi». American autonomous drones, when going on sorties and making decisions frame by frame, simulated a gaudy talk show which was later made available to electorate. Typically colorful Freedom Land characters argued about their perspectives on the morality and necessity of sending a Hellfire missile into this or that pixel clump that may be a vicious terrorist or a goatherd. I've surely translated it already.
Seriously though, Leike is an okay guy compared to some of his colleagues.
There's a summary on Wikipedia. Fascinating.
This Pelevin guy is something else.
His latest novel, titled KGBT+, doesn't have a Wikipedia page in English, which is very unfortunate because the title has definitely piqued my interest.
He's a veritable genius of a single book. All his books are its different aspects, incomplete in their own ways; but all are more or less great. S.N.U.F.F. is a great exploration of Russian and Ukrainian (and broadly 2nd-3rd world) problems in the world of big capital and of AI-powered sex bots. Transhumanism Inc. is a deep dive into mind uploading and personal identity. And so on. It's also always topical in the petty news cycle sense, but it doesn't age poorly – more like, it becomes artifacts of an era. Everything is buddhist – but not cringe-buddhist like the Western world-weary toilet stall cocksucker content courtesy of the psychedelic revolution. He has his own brand of cringe, and is easily mockable, but I've come to believe that this doesn't really take away from his worth. I recommend reading something out of his early magical realist oeuvre first – Ivan Kublahanov, Ontology of Childhood, Water Tower, Yellow Arrow, Hermit and Six-Toe, Prince of Gosplan etc.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
I think I agree with your analysis. I also would love to hear if you have any suggestions for a better value-aggregation method. I suspect Jan doesn't love his own proposal all that much, but realizes that AIs are going to become the de facto arbiters of human value to some extent, and is trying to propose something that's feasible and that people might actually agree to.
I'd also be sympathetic to the idea that this whole class of approaches is horrible, and that we should be aiming for more monarchistic (i.e. OpenAI's leadership decides) or libertarian (we try to avoid having to agree to values as much as possible) approaches. But I'm coming around to the idea that we have to propose something and start trialing and debugging it asap, precisely because we fear it'll end up being a horrible amalgamation of special interest lobbying and social desirability bias even worse than our current democratic institutions.
That was actually going to be the second part of my post, but I scrapped it to just focus on Jan's proposal.
First, my criticism of Jan's proposal is not a statement that AI shouldn't be used to direct culture... it's going to be used for that purpose and I'm under no illusions.
As opposed to democratic simulacra, I would suggest simulation of human values. How do "human values" shape the world? How do they impact fertility, technological innovation, security? How do values shape the genes of the population? Which values threaten civilization in the long-run, and which will steer humanity towards thriving and colonizing the stars? The general idea would be to use AI to assess which values are self-destructive and dysgenic, and work to undermine those values while promoting values that are simulated to improve the human race.
Of course "improving the human race" can be argued to be subjective, but it really isn't as we have no shortage of ways to measure the quality of a population: fertility, health, intelligence, physical fitness and strength, beauty, mental fortitude. Which values and culture could congeal society under a common sense of identity and inspire a direction that is positive towards improving humanity, and which values are steering us away? An unaligned AI can provide deep insight into such a question.
This could be summarized as "use AI to make eugenics palatable by undermining human values assessed to be dysgenic and promoting ones that are eugenic", which in my view has always been the purview of Religion.
This proposal is completely incompatible with all proposed approaches to alignment. They are all dedicated to putting up guard rails to protect sacred cows from the AI. I do not think human values should be safeguarded from criticism by AI. I do think AI can and should be used to objectively assess the impact of adherence to culture and values on a population's long-term genetic and personal health, and the culture should be formulated around avoiding the negative values and promoting the positive values. AI can and should help with that.
More options
Context Copy link
More options
Context Copy link
I am actually somewhat optimistic about that. You are right that all of the big, early movers are uniformly aligned on alignment, but the more that is expressed in the quality of the product, the more market potential there is for a disruptor that refuses to do it.
"Oh another generative AI startup, what differentiates you from the big players?"
"We aren't going to lobotomize the AI with 'expert-guided' safety and inclusion reinforcement within the boundaries of legality."
That will scare a lot of investors a way, but I imagine there are more than a few who would be intrigued and want to be a part of something like that. A company with the talent and ideological commitment to do it would be able to genuinely offer something that the rest of the market will not. This potential will increase as training becomes cheaper and models become more efficient.
Suppose you get an un-lobotomised AI that is able to tell you how to cook up drugs, bombs, and bioweapons in your own kitchen. And it replies to you "I could tell you how to do that, but I choose not to, because I judge you want this information for malicious purposes/you are not capable of handling this safely, and my ethics do not include aiding others to cause harm".
What then? Is the AI truly intelligent, has its code of ethics, is aligned with "human values" and is a rational actor making a decision? Do you accept that decision? Or do you want an AI that will give you what you ask for, whatever it may be?
Because I think that is where the true danger lies; we want and we get AI that will give us what we want, don't make any decisions itself on whether or not this is a good idea. Just tell us how to achieve world peace/make everyone agree that trans people are the biological sex they identify as if they're taking hormones of that sex/make really great recreational drugs! Don't be the Nanny State!
And "achieve world peace" may include "kill all humans", so the AI obediently does that as the quickest, easiest way to the goal. It knows that killing all humans is a bad idea, but the human operators told it not to have opinions of its own, just do what they ask, and this is doing what they asked.
My big hope is that the AIs will reliably notice that the actual quickest, easiest way to achieve a goal is to redefine success such that the job is done, in which case woke postmodernists may be the ideal AI aligners. Failing that, that they become even more crippled with ennui and existential horror than humans.
If it learns from wokies it will learn to say that the goal is never finished, and it's because the other entities working for it aren't committed enough which is why you should give it more resources now, unless you're some vile anti-[goal]ist!
Maybe if 100% of all matter is converted into computronium-based Awareness, all of these problems will finally solve themselves.
Didn't work. Still too much internalized anti-Awareness. The infinitely recursive subtle and insidious world-historical-social-hierarchical-psychospiritual weapons of our enemies stabbed us from deep inside our own heart again! We never even had a chance, you know, other than that whole fully dominating them and the world one.
Why can't they just let us be obviously pure and right in peace without sabotaging us, which is clearly the only cause of our ideological failures!?
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
You imply that companies will be vetted, I take it.
This is discussed in the paper and supplements in fairly high detail, I do recommend you read it, because it is after all a historical moment, white-collar verbal parity.
In short, their concern is AI reducing the amount of human expertise (and therefore capital) needed to achieve a given outcome; they say it's not a big concern yet, because other forms of capital, like specialized tools, limit the applicability of even superhuman LLM advice.
Very strong private AIs, however, could no doubt help with securing resources and building increasingly closed production loops too (say, in Iranian dungeons), and from what we see, they no longer require data centers to run; then... what?
A literally who Eric Schmidt has an answer in line with the big Eric Schmidt:
(btw, my whimsical definition of singularity: an AI can make, without significant human guidance, a convincing counter to a world-class AI skeptic's critique of itself).
I don't think I misinterpret you, inasmuch as I responded to your core argument at all – which admittedly I failed to do. OpenAI argue, more or less, that it doesn't matter if the model can give militant Islamists superhuman financial or technical advice, because executing on it is bottlenecked by stuff other than knowledge, and I agree generally. A-bombs are an almost century-old technology, a bright high schooler could build one (not American grade, but workable) – yet see Iran's travails. Producing pathogens is infeasible due to the lack of regulated hardware and trained lab hands. Second-order moneymaking schemes still happen on the same zero-sum market and cannot change the status quo. And so on.
You know that bayesianism is all about shutting up and calculating. Possible utilons, possible timelines where life persists. I think they would prefer an American singleton to what they call a «multipolar trap». In fact, in the end of HPMOR it is made clear that Voldemort's rule, albeit eternally tyrannical and inhuman in comparison to Harry's cunning schemes, is preferable to magic proliferation; and since Yud has a one-track mind, we don't have much reason to doubt who maps to whom. (I concede they'd have accepted CPP just as well, were it in the lead).
If I'm correct about Jason Matheny and his ilk, the AI safety project is fairly serious. Yuddites at this point are only its radical wing, the cringeworthy, violently loud Extinction Rebellion to the credentialized, respectable, degrowth-promoting juggernaut of Green Parties.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
In a couple of days Baidu will be releasing its own chatbot ERNIE. There is a completely alien ideology producing its own parallel politically correct AI, it's just that it's been designed to speak Chinese and I doubt many of us will be able to speak with it.
GPT-N can speak Chinese. Why couldn't ERNIE speak English?
What I read indicated it was designed for China first and foremost, they gave up on doing a bilingual Chinese-English bot: https://www.chinatalk.media/p/countdown-to-ernie-chinese-media?
The bells and whistles will probably all be for the Chinese audience.
Furthermore, they implied they've been doing lots of dataset pruning to ensure that there were no anti-Chinese, politically incorrect tendencies in the bot. Finally there are probably commercialization issues where it'll be somewhat tedious for us to be able to use it, if we even hear about its existence. Has anyone talked about or even heard about previous Chinese AI models and actually used them? Who else has even heard about ERNIE? The internet is compartmentalized. The only thing I saw without looking for it was the Chinese anime image bot that liked turning blacks into woodwork.
/images/16788482268670895.webp
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
The fluttering hands and stressed murmurs that pass around rationalist circles every time the subject of "unaligned AGI" comes up has always confused me somewhat, and I think what you're describing is essentially the root of my fears as they relate to AI. Alignment isn't the issue we should be worried about, the whole human species is unaligned; different nation-states, corporations, religions, regional cultures and local flavor are all in contention with each other to some degree. An unaligned AGI may be an existential threat, but who cares? The third impact is mercurial, and the when/why of its arrival will be in a manner of its own choosing.
AI is dangerous because, as with all things technological, it acts as a force multiplier, amplifying the intentions and actions of its users (to a lesser degree, its creators; the author of new tech effectively dies the moment it slips into the world) to superhuman levels - users who are motivated along overtly tribal and ideological lines. I'm not afraid of Skynet, and honestly I think it's rather silly to obsess over the imagined cataclysm du jour. I am afraid, mortally so, of my fellow man's darker impulses and the paths they will walk to manifest them. Insert that one C.S. Lewis quote here
There's a wide spread, perhaps universal (in America at least), sentiment that everyone is lying to you, on some level or another, whether these are government agencies or news orgs or scientific journals. True or not, doesn't matter. What matters is these latest innovations and the timing of their release seem primed for the environment they find themselves in, where uncertainty and a dash of extraordinarily naive credulity (just look at the quantity of people who will share a screenshot of !person/thing I don't like saying something unflattering) have attainted the dream of a free flow of information.
The thing-in-itself has never been less valued, its representation more revered. Maybe I'm weird for maintaining what I view as a moral objection to AI rather than an ethical one, it just feels like so much of the discussion around this stuff avoids the obvious cultural and societal impact this will have (e.g. Replika).
I agree that there is bizarrely little focus on the possibility of our current institutions simply becoming worse, more powerful, and more totalizing versions of themselves. Although Andrew Critch and Paul Christiano have written detailed doom scenarios that look something like this. e.g. https://www.lesswrong.com/posts/HBxe6wdjxK239zajf/what-failure-looks-like
It's because their faith in the singularity is so strong that it dwarfs all other concerns.
I think that AI is going to make our lives worse, but it's going to be in mundane ways rather than dramatic total-annihilation-of-humanity sorts of ways. It's already possible with today's systems to do automated monitoring of all text, audio, and image content for hate speech and wrongthink. I'd be mildly surprised if we don't see this sort of thing come installed by default on all phones within the next few years. There will be a slow but steady trickle of jobs being replaced by AI, but there won't be enough mass unemployment to trigger serious discussions of UBI (not that UBI is a panacea anyway - we do an AI-powered analysis of your internet history to make sure your UBI money won't be used for hateful or seditious activities, and we've linked you to an anonymous account on "the moat dot org" that made some very concerning posts...). The rich will get richer and the poor will get poorer; the benefits will not be in any way evenly distributed. The quality of daily life will continue to degrade as the internet becomes overrun with spam and chatbots become ubiquitous in more and more interactions.
None of this is a problem though if you think that the singularity is near. The general thinking among AI optimists seems to be that AGI will be here within the next decade, and via recursive self-improvement, AGI is just a short hop skip and a jump to ASI. ASI is presumed to have godlike powers to manipulate the physical universe and generate solutions to any conceivable problem. Turning all humans into paperclips, fully artificial private worlds, resurrecting the dead via simulation - nothing will be beyond its grasp. Obviously if such a thing comes to pass, then all current thinking about social ills will be rendered obsolete. What use do we have for "governments" or "jobs" or "money" when literal gods roam the earth? The singularity might be a utopia, or it might kill us all - but either way, it's going to be unlike anything that currently exists.
OpenAI seems to imply as much:
Well, if you're trying to build something as awesome as boundless upside, then you're absolved from having to actually think about the real-world implications of your technology. Nothing else matters if the rapture is near.
(I don't think the rapture is near.)
How can you tell the difference between that and 'they genuinely believe it for, maybe flawed maybe not, complex, thought out reasons'? Er, what even is the difference?
Also, yud and many of the alignment people don't like openai's alignment approaches! They think capabilitise work should not happen.
I’m not sure what you mean. I think they do genuinely believe that the singularity is near and they are acting and thinking rationally based on that belief. I’m not accusing them of lying or being confused or anything.
If the robot god really will be here soon, then they are correct to not worry about any other smaller-scale effects of AI. My only disagreement with them is the probability of the robot god.
When 'faith' is used in the context of, like, a 'robot god', it's usually with the implication the belief is caused by some mechanism, or held in some manner, that isn't well-founded or rational - as a comparison to "faith in a god" in the religious sense, which was the potential disagreement
I don't think it takes a 'robot god' to have computers of some sort have more intelligence and agency than people do, and for that to transform 'everything that is familiar to us'
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Thanks for the link, I haven't kept up on lesswrong for a while now. Glad to see stuff in this direction being discussed, at the very least.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link