site banner

Culture War Roundup for the week of April 10, 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.

14
Jump in the discussion.

No email address required.

Inferential Distance part 2 of ?: Minsky's Marvelous Minutia, or why I'm bearish on GPT

This post is a continuation of / follow up to my post on Inferential distance from a month ago, inspired by the recent discussions of GPT-4 and @ymeskhout's comments on prosecutorial immunity. I also feel like this might end up turning into a series, hense the "part 2" and the question mark.

Two things that came up in that previous conversation were a) the apparent differences between thing-manipulators and symbol-manipulators. That is people's whose job, hobbies, day-to-day life revolve around manipulating objects and those whose lives revolve around manipulating symbols/feelings. And b) the question of what constitutes a "hard" social problem, and how/why thing-manipulators and symbol-manipulators seem to have such wildly diverging opinions on that question.

For a bit of context my degree is in math but entering the field as I did, later in life having already spent 12 years in another career, I tended towards the more applied/practical side of the discipline. This tendency seemed put me at odds with a lot of my instructors and fellow students, especially the "nerdier" sort. That is those who were "nerdy" even by the relative high standards of nerdiness expected from someone pursuing an advanced degree in mathematics. for whatever reason showing an interest in applications was kind of looked down upon. To be fair, I did understand where they were coming from. From a young age we're trained to admire the brilliance of guys like Pythagoras, Leibnitz, Newton, Euler, Keppler, Einstein, Et Al. Afterall, why does anyone even bother to study math if not to follow in those men's footsteps and unlock the grand fundamental truths of the universe? In contrast, while the principals of kinematics, control laws, and signal processing, may be mathematically intensive they also come across as very pedestrian. Pure math guys seem to regard them with a sort of casual disdain, the sort of thing you delegate to unpaid interns and teachers' assistants. Meanwhile truth is you can build yourself a pretty good career working on control laws and signal processing, just not in academia.

This brings us to the question of what constitutes a hard problem. If you spend enough time working in robotics or signal-processing, you'll eventually come across Moravec's Paradox. The paradox is best summed up by this xkcd comic from 2014, specifically the alt-text which reads...

In the 60s, Marvin Minsky assigned a couple of undergrads to spend the summer programming a computer to use a camera to identify objects in a scene. He figured they'd have the problem solved by the end of the summer. Half a century later, we're still working on it.

...the "paradox" being that many functions that we consider baseline, and accordingly take for granted, are in fact extremely complex and computationally intensive. Whereas much of what we might label "higher reason" is actually quite simple and requires very little in terms of memory or processing power.

It turns out that it's relatively easy to teach a computer to play chess better than a human or to come up with mathematical proofs that are both novel and correct. And yet, after 60 years, despite the truly massive advances in both hardware and software represented by projects like stable diffusion Minsky's problem remains far from solved. In practice, you can pretty much graph straight line between the simpler a task seems/earlier a it appears in the evolutionary enviroment, to how hard it will be to replicate. Playing chess is easy, Bipedal locomotion is difficult. Bipedal locomotion only seems easy to creatures like you and me because we've been doing it since we were two-years-old, and our ancestors spent millions of years refining the techniques and bio-mechanics that were bequeathed to us as infants.

What does this have to do with anything? My answer is that I feel like a recognition/understanding of Moravec's Paradox is one of the major components of inferential distance between myself and most others both in the rationalist movement, and in academia. It is why I am reflexively skeptical of grand unified social/political theories. and It is also why I remain deeply skeptical of GPT and the oncoming AI apocalypse it allegedly represents.

One claim you'll see guys like Elizer Yudkowsky, Bryan Caplan, and posters here on TheMotte make on a semi-regular basis is that "GPT knows how to play Chess". But if you press them on the topic, or actually look at chess games that GPT has played it becomes readily apparent that GPT makes a lot of stupid and occasionally outright illegal moves (eg moving rooks diagonally, attacking it's own pieces, etc...). What this demonstrates is that GPT does not "know how to play chess" at all. At least not in the same sense that Deep Blue or my 9-year-old can be described as "knowing how to play chess", or AlphaGo can be described as "knowing how to play Go".

Furthermore, once you start digging into their inner workings this lack of "knowing" appears to be a fundamental weakness of the Large Language Model architecture. At the end of the day it's still just a regression calculating the next most plausible word (or in the case of GPT-4 string of words) based on the correlations found in it's training data. Granted GPT-4 is certainly a step up from GPT-3 in terms being able to pass as human. The shift towards correlating longer statements rather than individual words seems to have plastered over a lot of the jarring discontinuities that made GPT-3 generated posts so easy to pick out. In contrast GPT-4 can actually kind of pass for human from the proverbial 50 ft away. Unlike prior GPT iterations, identifying it actually requires a level of careful reading or some sort of interaction.

Eugene Volokh's posts on Large Libel Models probably deserves a discussion of their own but INAL and not really interested in questions of liability. In any case he ends up running into the same issue with GPT that I did. Users here talk about instances of GPT "lying" or "hallucinating" and how to reduce the frequency of such instances, but the conversations inevitably devolve into self-referential nonsense because neither of these terms really describe what is actually happening. In order to "hallucinate" one must first be able to perceive. In order to "lie" one must first understand the difference between true and false. and GPT possesses neither. Simple fact is ask GPT for five examples of prosecutorial misconduct complete with citations and newspaper quotes and it will provide the names of five prosecutors, their alleged crimes, some juicy quotes, and supposed case numbers. However while the names provided might actually be real prosecutors, and the media outlet quoted might be a real outlet, if you actually look up the court records or try to find the quotes you're going to come up short because the example was not something that was pulled out of memory and provided, it was "generated" form the prompt in exactly the manner that a Large Language Model is designed to do.

to be continued...

edit: fixed link

Furthermore, once you start digging into their inner workings this lack of "knowing" appears to be a fundamental weakness of the Large Language Model architecture. At the end of the day it's still just a regression calculating the next most plausible word (or in the case of GPT-4 string of words) based on the correlations found in it's training data.

This is something that I find very unconvincing on the anti-AI side of the debate. First one is what I will call "just" argument. GPT is just next word prediction machine, it is just stochastic parrot and so forth. This type of arguments seem to argue that certain method such as training LLMs on predicting text will obviously result just in text predicting system. Which I think is red herring - training on text is obviously sufficient for LLMs to develop qualitatively different capacities such as multimodality. As the old saying goes - quantity has quality of it's own. It seems to me that it should be on proponents of just argument - who pretend to have deep knowledge of these models - to explain and possibly predict these things before saying stochastic parrot .... and therefore multimodality. Plus of course these types of cheap arguments can be used against humans - human brain is just a product of blind evolution. Or as in this book review, human brain is just a multi-layer prediction machine.

It seems to me that for AI the focus is a lot on hardware, the training process or on the output. But for humans it is always highly spiritual focus on qualia, feeling of understanding and other subjective things - it is not about brain structure, or pointing out how humans produce stupid output and therefore brains cannot truly understand, they do not know, they do not have representation of the universe or that they cannot plan. There are more obnoxious types like this episode of Adam Ruins Everything but there are also other and more sophisticated critics - the common denominator of all of these is that they are awfully certain to know what is [not]happening inside LLM models. I do not see many legibility experts who would really claim to know for certain that LLMs do not understand. Because who knows what is happening in this whole mess of inscrutable matrices of parameters, maybe somewhere in there is some kind of representation of the universe. We certainly do not know what is happening inside human brain when we scan it - unless of course we use Machine Learning for that But more importantly, LLMs can predict text on par of some of the top percentiles of people who understand, know or plan. So yes, LLMs can pass test specifically designed to test for understanding, they can produce plans on par with human planners and so forth, but for some reason despite all that one can simply claim is that they do not truly know or plan because of stochastic parrot or some such.

More convincing argument - or The Motte if you wish - is that LLMs do not understand, plan etc. like humans. Which is perfectly reasonable argument, except that they do kind of develop certain things that humans and also some animals also develop. So they are like humans in certain way but completely alien in other ways. However even this is loaded questions as LLMs can produce some output equivalent to humans but they may still not do it like humans. But each new implementation of these models are improving in certain tasks that were still outsourced to Mechanical Turks, the space for unique human application in this space is narrowing.

Now I have to say that I do not know where this all will lead. It may very well be so that current Transformer approach will reach certain plateau and then stops. There may be significant areas where humans will remain superior, and it may even have something to do with the fact that "Auto-Regressive LLMs are exponentially diverging diffusion processes" as LeCunn says. I do not know, but neither do these people. What I see is quite a rapid growth in capabilities of these models just with more compute.

I suppose here is as good a place as any to drop my two cents:

I think one of the things that definitely makes AI more machine than man (for now) is something that I assume is fundamental to "consciousness:" motivation. What we call "agency" is possibly confused with the word "motive." As conscious beings, we humans have the following: a sense of self, some un/subconscious schema of instincts, and motivation. We are things, we can do things, and importantly, we want to do things. The mystery stuff of "qualia" that IGI argues for above is something we don't perfectly understand yet--is it just a biological form of training/pre-tuning written into our genetic code? Is there something spooky and supernatural going on? Is there truly something that makes us different from all the animals that can't build anything more complex than a nest, dam, or hidey-hole, something other than just a bigger brain?

Currently, GPT is a mechanical thing that won't do anything on its own without being fed an input. This is probably why anti-doomers take the "just unplug the AI 4Head" stance: to them, the AI lacks an innate drive to do anything it hasn't been told to do. If GPT is a baby, it's a baby that will just sit still and make no noise.

Maybe this is the real crux of our current moment: while these AI models are plenty capable, some just can't make that leap to "these are just like us, panic" because we aren't having to practice yomi against a motivated entity.

The mystery stuff of "qualia" that IGI argues for above is something we don't perfectly understand yet--is it just a biological form of training/pre-tuning written into our genetic code? Is there something spooky and supernatural going on? Is there truly something that makes us different from all the animals that can't build anything more complex than a nest, dam, or hidey-hole, something other than just a bigger brain?

A lot of people come to this class of arguments, humans are somewhat unique as they posses agency or motivation or qualia or in the past it was creativity and so on. It reminds me of the famous Chinese room argument where Searle smuggled in the concept of "understanding" by inserting literal human into the thought experiment. If human does not "know" Chinese, then the system itself does not know it either, right?. This is our intuition about knowing - mechanical systems cannot "know", only humans do and the only human around in this thought experiment does not know, QED. The most straightforward criticism is that human does not represent any cognitive agent in the whole room, he is just one part of the algorithm of making output. The room as a system can be capable of "understanding" on its own. And yet this whole argument is used over and over and I see something similar now with AI. As I argued above, people are all too ready to describe AI systems as pieces of hardware, as a training mechanism and so forth, they do the utmost to "dehumanize" AI with all these just arguments. And on the other hand they are all too ready to describe humans only subjectively, as agents possessing qualia and understanding and with capacity for love and creativity and all that to maximally humanize them. They never mention brain or how human neural network is trained or how cognitive algorithms work, no it is all about wonderful internal experience so unique to humans and so unlike just machines.

I really like a quote from Yudkowsky's essay How Algorithm Feels From Inside

Before you can question your intuitions, you have to realize that what your mind's eye is looking at is an intuition—some cognitive algorithm, as seen from the inside—rather than a direct perception of the Way Things Really Are.

People cling to their intuitions, I think, not so much because they believe their cognitive algorithms are perfectly reliable, but because they can't see their intuitions as the way their cognitive algorithms happen to look from the inside.

I think this is about right. For all we know before LLMs make an output they may have some representation of what is "correct" and what is "incorrect" output somewhere in there. As argued before, LLMs can spontaneously develop completely unique capabilities like multimodality or theory of mind, it may very well be so that something akin to subjective feeling is another instrumental property that can appear for even more developed system - or maybe it already appeared but we will not know because we do not really know how to test for qualia.

But I still think it is all a red herring, even if LLMs will never be conscious and they will never be able to think like humans; we are currently beyond this question. It truly is immaterial, our current crop of LLMs do produce high quality output on par with humans and it is what matters. Really, we should drop this unproductive discussion, go and play with Bing Chat or GPT-4 and see for yourself how much good did all these qualia debates for you.

In a sense it is even more scary that they can do it without developing complete set of human-like properties, that fact bodes unwell for alignment efforts. To use an analogy, recently it was found that Alpha Go was beaten by a very stupid strategy. It seems that all the critics were correct: see, the neural network does not really understand Go, it could be fooled so easily, it is stupid and inferior to humans, it lacks certain quality of human mind yet. Now for me it was actually terrifying. Because for years Alpha Go was considered as a superb Go player beating the very best human players who dedicated their whole life to the game. And now after years we found out that it was capable of doing all that without even "knowing" what is was supposed to do. It obviously learned something, and that something was sufficient to beat the best humans for years before the flaw was spotted.

It is incredible and terrifying at the same time and it is harbinger of what is to come. Yeah, GPT-5 or some future system may never have qualia and agency and that special human je ne sais quoi - but it will still beat your ass. So who is the sucker in the end?

[Ramble incoming]

I guess, then, between the Chinese Room and AlphaGo and AI art and GPT, what we're really worried about is meaning. Did AlphaGo mean to be so good? What does it say when it rose to the top and the damn thing doesn't even "know" in any meaningful way what it did?

Kind of calls back to the recent thread about the Parable of the Hand Axe. For most of human history, our works were judged not merely by the output, but the journey. We appreciate the artist's processes, the engineer's struggles, the scientist's challenges, the warlord's sacrifices, the king's rationales, and so on. AI has recently provoked so much backlash because some realize, correctly or not, consciously or not, that AI threatens to shortcut the meaning imbued in the process of creation. Effortless generation of anything you want, but it will mean nothing because there's no "soul" to it.

I'm sympathetic to this argument, but I also have the capacity to acknowledge that maybe the way we think about "meaning through struggle" has the potential to become outmoded. On the third hand, though, it might be mind-boggling and embarrassing to think that humanity operated this way for so, so long. On the fourth hand, however, maybe the fact that current AIs were trained on the scraped works of a significant chunk of humanity does contain meaning in of itself--if meaning is achieved through the struggle and not the end result, AI still counts, just for the entire species and not merely the few.

I think meaning is another of these subjective/human concepts that may be useful but that are also dangerous, because it starts with the premise that humans are unique. But from other standpoint humans are "just" result of an evolutionary process that optimizes for inclusive genetic fitness. Imagine that we really are living in a simulation where somebody started the whole Life game by introducing Earth environment and simple rule for biosphere to optimize for inclusive genetic fitness. Except in a few billion ticks, the simulation produced species homo sapiens that evolved algorithm that can "hack" many instrumental goals that evolution developed as implementation of its main goal. One of those things for instance is sexual drive to increase number of offspring - humans were however able to hack this by being able to masturbate or use condoms. They sucked out the "meaning" of this activity, or maybe they found their own meaning there - to great exasperation of our simulation designer who now observes something strange happening in his model.

To expand the analogy, "optimize for inclusive genetic fitness" is akin to "optimize for predicting next word" in world of AI. Then goal of "learn to play Go" is akin to "have a lot of sex". But the Alpha Go somehow hacked its programing so to speak and learned something different, it decided to not play Go in a sense humans though it will. One can speculate that it developed its own meaning for the game Go and decided to stubbornly ignore whatever was meant by its creators. That is what I meant about bad news for aligning, whatever the LLM learns can be absolutely orthogonal to system used to train it (be it darwinian evolutionary process or next word prediction for a text) and it can be orthogonal even to some very detailed observation of output that however is superficial under many conditions (such as homo sapiens shagging like rabbits or Alpha Go beating good human Go players for years). What happens under the hood can be very hard to understand, but it does not mean it has no meaning.