The first time I saw an AI creating images in like 2018 or so, I remembered the scientists called it "dreams" and I thought it was so appropriate. The constantly morphing landscape and how things blend together to form new things and how almost everything is wrong when look at closely. It was an interesting description, and I didn't care much more than that.
Then the AI got better, and still the dreamlike things never go away. The constant problem with the hands, words never properly formed, clocks always look wrong, etc. But now it is more cohesive, more "solid" but still malleable. Like a lucid dream. Still straddling the border between consciousness and unconsciousness but there is a hint of control and direction now.
As someone who lucid dream, I think the AI images and the LLMs are just like a person dreaming right now. They sort of have control, but actually not really. Kinda hard to explain but even when I lucid dream, I know it is a dream and I can bend it to my will but at the same time, it isn't possible to control my thoughts. Trying too hard to assert control and I wake up. So it is still a state of unconsciousness for me and not at all comparable to the "me" when I am fully awake.
Of course, the AIs can't wake up if we use that analogy. They are not capable of anything more than this state right now. But to me, lucid dreaming is already a step above the total unconsciousness of just dreaming, or just nothing at all. And wakefulness always follows shortly after I lucid dream.
A more conservative, but less entertaining and less poetic model is:
You lucid dream when your "explainers" and "predictors" are operating on the random feedback and subtle noise that's present while your mind is defragging and moving data from short term to long term memory.
Random noise, with low signal, fed into something that is trained to fill in details kinda explains how Gen AI works, up to a decent approximation.
Your mind has a lot of these predictors and detail fillers. All those are trained on your experience. I think dreaming is either tweaking those predictors and you get to watch, or it's in fact a junk signal you're not supposed to be able to see, but you happen to be kinda awake enough to form memories from it.
Using words like "wake up" imply deeper and darker depths to these networks that I doubt exist.
Descriptions of psychological phenomena organized around made up nouns are not necessarily more grounded in reality than ones that make creative use of existing verbs.
You'll notice we're all in lala land here. You can have a meaningful conversation at a speculative level.
But the nouns are analogies for regular old predictors and detail fillers we know exist. For examples, point to your visual blind spot quick. Or go watch your eyes move with a mirror. Or do any of the really neat experiments in the Coursera course "Buddhism and modern psychology".
I'm extrapolating, and yeah, landing right in the middle of "who knows".
I feel like lucid dreaming is in a way the baseline and that when we're awake, reality constantly knocks us back into groundedness. In a lucid dream there is no error correcting signal to keep our fabulations in check. Things tend to drift away like the silhouettes of an infinity mirror. Reality is like the flash of solid pixels when switching away from the hysteresis stream in a shared window Zoom call.
As someone who lucid dreams frequently, and has done a good deal of experimentation in that state, the fundamental difference between your sleeping state and your awake state is that some of your brain is just 'off'.
The parts of your brain that are on do their best to match your experiences (or just replay experience), but for things that concern the off parts (counting, reading, reasoning...) those elements are just missing. When you look at a clock it doesn't look like nonsense, it doesn't look like anything, as the part of your brain that interprets that information, and would therefore be necessary to simulate it, just isn't present. You 'see' all of the visual qualities of a clock kind of jumbled together the same way a lot of generative AI produces them.
The feeling I get from messing around in those states a lot is that there is no 'baseline', just some things that are on, and some that are off. When you dream, it's the visual cortex that is on, so things concerning light and shape and texture exist and you experience a slimmed down consciousness composed of those pieces. When memory turns on, you experience references to other things, and chain thoughts together (I think this is what separates normal and lucid dreams).
So far, those are the only two things I can turn on without waking up. Anything that uses reason and bam I'm awake. If I count, read, or recognize inconsistency, it's over.
Consistency is possible, but only forward facing. Like you can string together thoughts and events with purpose, and a lot of them will be consistent, but the moment you look backward and try to ask "did y follow x" the lights come on.
> but for things that concern the off parts (counting, reading, reasoning...) those elements are just missing
> Anything that uses reason and bam I'm awake.
Tangent, I can reliably wake up from nightmares / bad dreams by going "oh that doesn't make any sense, that can't be". There's a moment of the bad dream "trying to adjust" but then it feels like something just gave up.
Wow, absence of error correction explains why dreams are the way they are... No wonder my dreams always start out normal and mundane, but end up with extremely exotic events in the end (even though I don't realize it during the dream).
I did notice that AI 'art' basically got less interesting as it got 'better' (read: higher fidelity). The collages of conceptual spaces in the article are infinitely more interesting than any of the prompt-driven attempts from the popular models. They actually communicate something about the thing they're depicting by showing the ways in which the model deviates from it. That's a real artistic expression.
Less appealing and useful, absolutely no. Look at how the midjourney sub grew. It was ultra niche when it was v3 'dream scenary'. Its mega popular with v5 'ultra-realism'. Art involves technical execution, significantly so, not just 'artistic expression'.
The explosion in popularity of AI imagery isn't because the results became more artistic, but because they made it easier for people to mimic technical execution without understanding or caring about making art.
Put differently, it made it possible for them to have the technical execution without being able to understand or change why certain things are appealing to them. That's why most of it looks the same and why the vast majority of the results don't have any appeal beyond satisfying the ego of the person who generated it.
In a way, the appeal of 'modern' AI generated imagery to its users is similar to someone who thinks chess is cool, doesn't care enough to understand it, but wants to play anyway, so they use a mediocre chess engine to generate their moves against other players and rely on playing a massive number of games to make their win count look good.
> The explosion in popularity of AI imagery isn't because the results became more artistic, but because they made it easier for people to mimic technical execution without understanding or caring about making art.
It's almost like when people first started buying really fancy DSLRs with no other training, and took a lot of magazine-quality amateur photos.
SD raises the bar a little though-- it trained on a large body of professional images, so generating anything non-fantastic is probably going to be framed and composed half-decently.
...which is the problem. It's all too perfect. Wabi-sabi is dead; everybody can impersonate a professional artist/photographer now, so everything they're producing is the equivalent of photorealistic, award-winning, 8k motel art [...by Greg Rutkowski].
I disagree on the "all too perfect" part. It's pretty much exactly like when people first started buying fancy DSLRs and so they took a lot of photos that were technically great but weren't interesting content wise.
The images generated by SD etc are technically great, but the content is all 'same-y' with no individuality, to the point that people used to seeing images of that kind of art can very easily recognize when they're looking at AI generated images, even if they don't have any of the telltale technical errors (like poorly defined fingers).
The difference is just that getting a fancy DSLR was still on the path to potentially developing photography skills, while using image generators isn't. So while those interested enough in photography might've eventually developed their skills further, those using image generators just become abusive and disrespectful towards actual artists for not 'acknowledging' them as artists.
It's perfect until you want something specific, then it's a pain in the ass. Even with LoRA models and ControlNet and everything else, the lack of any essential control is always going to be limiting.
To play devil's advocate though, there's nothing wrong with creating something just to satisfy your own tastes. The problem comes with trying to pass yourself off as a professional artist using these tools.
Yeah, I agree with that, nothing wrong with people generating images for their own satisfaction and even sharing them as long as they're clearly tagged as AI generated. Just like how computer assisted chess is allowed to be its own competitive thing called 'Advanced Chess'.
However, just yesterday I came across someone trying really hard to play at being a pro artist, trying to show recordings of their drawing and painting process when they were obviously using AI (by roughly tracing over the AI image, masking the AI image underneath, then recording themselves erasing the mask, revealing the AI image underneath).
In certain artistic sub-fields I'm interested in, like with anime art, I've seen a lot of vitriol thrown around by AI 'artists' because they haven't been through the learning process through which they could empathize with actual artists and the basic courtesies that exist in the community. This has been largely responsible for my negative opinion on the social value of recent developments in AI generated content.
To give some examples, there was an incident where an artist had been streaming themselves drawing something, only for someone to take the incomplete image from the stream, have an AI 'finish' it and post it as their own before the stream even finished. Then there was a more serious incident where someone finetuned a model on a specific popular artist's work and started to sell generated images under the original artist's name. This one ended with the site hosting the content having to change their rules to ban monetization of AI art.
I think it's a stretch to call Midjourney output art and technical execution. I think it's more "content" and "regurgitation". And most, not all, of what trends on the subreddit is pop culture mashups. It grew because people love Star Wars and Disney and Game of Thrones and so on and so forth.
Further instrumentalization of art to produce a median state of being appealing is IMO not a laudable goal. Everything Adorno predicted about the general course of the culture industry in Capitalist economies was more or less correct and AI art is just accelerating these trends.
Map this same logic to how we use language in creative works: would it ever be desirable to reduce all of language’s diversity and variability to a similar state? How is that distinguishable from the kind of Orwellian horrors that Americans love to decry?
The dream-like appearance of AI generations is really interesting. Humans, when creating art, use our representations for objects to control muscles to move a brush or pencil across (digital) paper. What we see from AI is what you get when you remove the "muscle module", and directly apply the representations onto the paper. There's no considering of how to fill in a pixel; there's just a filling of the pixel directly from the latent space.
It's intriguing. Also makes me wonder if we need to add a module in between the representational output and the pixel output. Something that mimics how we actually use a brush. How we consider what to paint where.
>Then the AI got better, and still the dreamlike things never go away. The constant problem with the hands, words never properly formed, clocks always look wrong, etc.
Might be a bit behind the times, Midjourney seems to have completely solved the hand issue from what I can tell.
As I see it, Stable Diffusion is a camera you can tune to take pictures from alternate realities. The images it creates are of real people and subjects, just in some parallel universes.
Photo-surrealism is one of the most interesting facets of AI artwork in my opinion, and the closest thing to legitimate artistic expression since by definition you're trying not to replicate something. It's a shame all anyone wants to do is anime porn.
> we can progressively increase the numerical values of the weights—eventually in some sense “blowing the mind” of the network (and going a bit “psychedelic” in the process)
I wonder if there's a more exact analog of the action of psychedelics on the brain that could be performed on generative models?
Andrew Gallimore's online course on psychedelic neurochemistry[1] argues that psychedelics disrupt something like a constant fine-tuning cycle taking place in normal perceptual activity, causing new activations between cortical structures where connections had previously been dormant (see [2] for an in-at-the-deep-end account), causing the generative model in the higher layers of the cortex[3] to adjust in an effort to regain predictive efficacy.
I wonder if some such kind of "dynamic" generative mechanism might be needed to perturb an artificial model analogously?
In fact I didn't realise it to start with, but after I'd scrolled for half an eternity, and saw there was still an eternity of content to go (I guess that's how eternity works), I thought 'huh, it's really long, is this a Stephen Wolfram piece?'. I don't know how he does it, he must type in his sleep.
Keep in mind that this is all about diffusion models.
The latent space of GANs is known to be structured differently - much smoother and more disentanglement of concepts. Very likely each type of model has a unique latent space.
Yes, you wouldn't get anything like what he shows in all these pictures if you were walking a BigGAN or StyleGAN latent space rather than the 'latent space' of Stable Diffusion 1 (which isn't really a latent space to begin with). So that rather undermines any connections he might want to draw to grander things than just SD1.
There is a logic jump in the concept of the ruliad.
First it goes:
> the entangled limit of everything that is computationally possible: the result of following all possible computational rules in all possible ways
And then
> it encapsulates not only all formal possibilities but also everything about our physical universe
The assumption that everything about our physical universe is described by computational rules is not a given, and is never explored nor justified. Even the assumption that our perception follows computational rules is not really explored, just given.
There is this old idea that with an universal computer that knew everything, one could predict anything. But that has been challenged by uncertainty principles, observer effect and all that.
This is a well-known conjecture called the Physical Church-Turing Thesis.
It is just a conjecture, but emperically it has held true to the limits of our knowledge of physics. It also still works even if the universe is probabilistic.
> but emperically it has held true to the limits of our knowledge of physics
Considering that our entire knowledge of physics is encoded as computational rules, because that's all we really know how to do, this statement is basically a tautology.
No, we do physics this way because we don't know any other way. When this way fails (e.g. quantum gravity, turbulence), we assume that the problem is that we haven't tried hard enough yet. Which may be true, but it's also possible that the computational approach simply doesn't work for those phenomena.
This sounds an awful lot like "God of the gaps" argumentation with "non-Turing hypercomputation" substituting for "God". If the computational approach doesn't work — that's all of human reason, to emphasize: any symbols you can scribble on paper and reason about over coffee, that's a finite algorithmic process — then there's no humanly-achievable endeavor that can solve the problem. It's forever hiding in the gaps, the shadows; it's definitionally unknowable.
So what field of science are we discussing right now? One of the more frivolous branches of philosophy: talking about something that we can't define, can't reason about, won't ever be able to reason about; something which marks no experimental evidence in our human sciences. Therefore: it doesn't exist. Apply your favorite razor. Something we can't possibly talk about, let's not talk about. Something whose existence or non-existence are indistinguishable, which leaves no shadow in the observable universe: it does not exist.
It's not a "God of the gaps" argument because I'm not claiming that "God" (that is, incomputable physics) exists. I'm simply pointing out that the fact that our entire understanding of physics is built on a computational framework doesn't actually say anything about the Universe, merely about the way we do physics. There is a priori no reason to expect that the Universe would be in any way amenable to human understanding. We have just found that in some cases, it is.
And non-computability doesn't make something philosophical or mystical. We have encountered phenomena beyond the reach of any computation in mathematics and computer science many times. They exist, no matter which "razor" you apply.
- "We have encountered phenomena beyond the reach of any computation in mathematics and computer science many times. They exist, no matter which "razor" you apply."
The razor applies as to whether the physical universe runs on laws that don't fit in Turing machines. It applies because it's an unfalsifiable proposition. You can't embed a Turing machine in a super-Turing universe, ask it "what is the nature of your cosmos?", and have it determine (within finite computation) there are more things in heaven and earth than it can compute.
(This is the ordinary Church-Turing thesis: if some finite interaction between Turing observer and super-Turing universe convinces the Turing observer of the super-Turingness of its universe, within finite time... then an ordinary Turing machine can also do exactly the same thing, by simulating the first Turing machine, and brute-force enumerating every possible interaction within that finite bound. The nature of super-Turing machines is incomprehensible to Turing machines; there's nothing a super-Turing machine can do to prove it exists to a mere Turing machine).
you know... the entire point of conceptualizing a hyper-Turing machine in the first place, the key difference between a hyper-Turing machine and a Turing machine is the solvability of the Turing machine halting problem.
So shouldn't a Turing machine be able to determine if a hyper-Turing machine exists by presenting that very problem? I.E. problems that take O(n!||n^x) to solve but O(log(n)) or less to verify.
The Church-Turing thesis as you describe it seems to miss this idea, that things can be verified in a different time than it is solved. Is that your interpretation or as it is written?
It does not follow that if a hyper-Turing machine can convince a Turing machine of it's hyper-ness, then the non-hyper-Turing machine is actually hyper to begin with. For reasons stated above. A hyper-turning machine should be able to give proofs to problems verifiable by a turning machine, but that are not solvable within a given step-count/time-frame by that Turing machine.
The interactions between a hyper-Turing machine would be as follows.
```
T:"Yo, solve this!" - some O(n!) "for 10 million inputs"
HT:"done, is x"
T:checks notes "that's right, did you have that saved, that was very fast."
HT:"nope, from your perspective, I might be guessing the answers, you saw me take no steps"
T:"How do I know you aren't just guessing?"
HT:"I'm always right"
T:"How do I know you aren't just always guessing the right answer"
HT:"That's the neat part, you don't."
T:"So I'm going to make a detect-halting program, put it in itself and run it, will it halt"
HT:"It will do " x
detect-halting program does x.
T:"That's pretty strong evidence there... I could never say that"
```
I'm thinking and writing at the same time, so my final thoughts are... if a Hyper-Turing machine is inexplicable to a Turing machine, shouldn't the Turing machine observe the Hyper-Turing machine do inexplicable things, thus revealing itself to the Turing machine?
You're conflating computability with computational complexity; that seems to the source of most of your confusion.
I don't remember the name of the complexity-theory analog of the Church-Turing thesis; but it's an open problem what types of computers physics can support the existence of, how they relate to classical Turing machines in complexity theory. See for example: quantum complexity theory.
About halting: there is provably no way to verify an oracle for the Halting Problem, even if you had one. You can't even ask a hypercomputer to write a proof for you: in general there exist no finite-length proofs.
As someone who works in a custom fabrication shop and who has been research job shop scheduling for over a decade as an attempt to get better at it, it's a computationally impossible task. There are huge numbers of "optimal" ways to schedule the shop and all of them require enormous amounts of computation. Genetic algorithms are offering the most promise right now, but the schedule has to be constantly updated because things never take as long as you think they will.
This is primarily because, unlike manufacturing, the basic unit of our system is a person, with all of their chaotic outputs. You start stacking those up and you end up with "long, fat tails" in every analysis you do. Your probabilities are spread almost flat over a broad range of outcomes.
It's fascinating and definitely falls into the category of uncomputable.
I'm not clear what you mean about turbulence? Quantum gravity involves as-yet unknown physics but turbulence is surely just a case of "the number of calculations we need to simulate it grows quicker than we can reasonably keep up with"? i.e. simple toy simulations work fine but we don't have a planet-sized computer to do something more extensive.
(Just to clarify - this isn't a "classical vs quantum" thing - it's a "we know the equations vs we don't" thing. I'm sure QED simulations are fairly similar within a known domain)
Presumably the GP refers to turbulence being infamously difficult to model analytically. But we have a great model for turbulent flow, the Navier–Stokes differential
equations, it’s just that we can only solve them numerically in all but the simplest of cases. But when we do we get good results, so it’s not like turbulence is some mystical phenomenon beyond the reach of our standard mathematical tools!
The alternative is numerical approximation, as I mentioned. Like always in physics, any analytic (ie. "closed form") solutions would also necessarily have to be special-cases, approximations, and simplifications. The problem with turbulence is that it's so chaotic and complex that it doesn't seem to be reducible to simpler models while still retaining some predictive power. But that's not very surprising; we're talking about the chaotic motion of molecules at the Avogadro scale! After all, we can't even write closed-form solutions (or even approximations) to the motion of merely three bodies under gravity, never mind ten to the power of twenty-three.
But what is fascinating about turbulence, though, is the "edge of chaos" – the boundary conditions at which laminar (non-turbulent) flow suddenly turns turbulent. On one side of the boundary analytic treatment is possible, on the other it is not.
> It certainly hasn't conquered consciousness and qualia and subjective experience.
I'm not sure how you're so certain of this. It is entirely possible that all we need to do to create a conscious mind is run a full molecular dynamics simulation of a detailed enough human brain scan. We know the equations to do this, we just don't have nearly enough computational power.
That would only be possible if qualia has no causation. In the case where our experiences have real effects they cannot be fully described by a Turing architecture.
How are you certain? Has someone published a paper that I'm not aware of that settles this?
You are missing a dimension here by the way. Even if we could find a physical location of the brain that lights up when the color red is experienced, that doesn't have any explanatory power for the nature of subjective experience. How can it be possible to experience being a brain, being a brain itself? Not just observing another brain and correlating brain activity with reported experiences.
Experience and consciousness are axiomatic and precede theory about the physical world. You started with sensory experience and perception, and built models of how your brain works on top of that, not the other way around.
> How are you certain? Has someone published a paper that I'm not aware of that settles this?
That's kinda the point. I'm not certain at all, but given the available information, it is the simplest explanation.
> How can it be possible to experience being a brain, being a brain itself?
I'm not really sure what you're asking here. A clock can tell time, but if it was _like something_ to be a clock, it would feel like knowing what time it is, it would not (and pretty much couldn't) involve awareness of all that goes into producing that feeling (because then you'd need a mechanism for awareness of that, and then a mechanism for awareness of that mechanism, and etc). The difference with the brain is that we're pretty sure it is _like something_ to be a brain in a body, and we also know it doesn't include perfect awareness of what goes into producing that feeling. So when you ask 'how can it be possible to experience being a brain, being a brain itself' I would say that it isn't? because your awareness is limited to the qualia that bubbles up out of the mechanics in the brain, and your experience of being a brain is similar in fidelity to the clock's experience of being able to tell the time.
> Experience and consciousness are axiomatic and precede theory about the physical world. You started with sensory experience and perception, and built models of how your brain works on top of that, not the other way around.
This is just solipsism. Sure, it is entirely possible all of the experiences and perceptions I've experienced have misled me into thinking that the brain is all there is, but that seems kind of contrived? I'm just as confused as you about why we aren't all p-zombies, but, well... the history of all knowledge has basically been "yeah it turns out to just be physics", so my guess is that qualia is also just physics.
It's not solipsism. The idea that "you" exist is just as conceptual as physics - there is no solipsism without "you". And the idea that it "turns out to just be physics" is a concept rather than reality. It's a thought of yours, not ground truth. And labeling this "solipsism" is just falling into that trap once again. It's really a miasma of perception in the form light and shadow and color and sound and pressure and heat (which are crude conceptual labels themselves for actual experience), and thoughts emerge in this cloud of sensation, also as independent phenomena in themselves. This subjective experience is first and primary. Any structure you build on top of it is conceptual and not ground truth.
It's not solipsism or contrived, quite the opposite. It's profound. Whether we like the fact or not, or that it's not convenient to your desire to have a mathematical formulation of the world, doesn't make it any less true. To downgrade the most direct experience of reality and give primacy to the conceptual is to confuse the map with the territory.
> Experience and consciousness are axiomatic and precede theory about the physical world.
Actually, they seem unnecessary to any theory about the physical world. They likely don't exist, but if they do, it's just bad psychology that wasn't quite debilitating enough for evolutionary pressure to weed them out.
I don't even think I'm conscious myself in the way you imagine yourself to be. "Consciousness" probably belongs in poetry, it definitely doesn't belong in philosophy and that goes double for real science.
So just ignore first hand experience as a phenomenon at all? That seems like a huge thing to pull the wool over.
It absolutely belongs in philosophy, and is one of the core topics in any philosophy program, and I'm not talking about "real science", which is precisely why I said that physics doesn't explain it.
There's nothing to ignore. Some evolved monkeys have vivid imagination, which they mistake for reality. Phenomenon noted, time to move on to important things.
Are you speaking as the representative of "reality"? Are you somehow the embodiment of "science"? Is it possible to experience and reason about reality without going through the eyes of an "evolved monkey"? You are speaking as if you are disembodied and floating around in the ether experiencing reality directly. It's so close to you that you don't even notice that you exist, and you are the only window you have to reality. There isn't any reality that you can be aware of outside of your own subjective perception, despite you "vivid imagination". You are confused if you think "reality" and "your model of reality" are two different things.
Already you've sneaked in the presupposition that consciousness is only a result of your brain activity. Have you ever not eaten for two days? Or not slept for more than 24 hours? Have you ever had an itch under a cast that you couldn't itch, and been constantly distracted by it? You think that doesn't affect your consciousness?
> Already you've sneaked in the presupposition that consciousness is only a result of your brain activity.
What else could it be? The interesting question is how and why.
> Have you ever not eaten for two days? Or not slept for more than 24 hours? Have you ever had an itch under a cast that you couldn't itch, and been constantly distracted by it?
Sure, I've done all of those things, but they all cause physical changes in the brain. These physical changes then directly affect your qualia-type experience of reality.
Have you read "What is it like to be a bat?" by Thomas Nagel? It's readily available online as a pdf. It discusses the hard problem of consciousness and the mind-body problem that I only briefly and poorly hinted at with my earlier questions. Reductionism doesn't work for consciousness. At all.
It hasn’t conquered beauty and comedy and love either. The things you mention will turn out to be folk human notions without satisfying scientific descriptions. They are no less real for it, but we simply don’t ask physics to describe love. I suspect consciousness will turn out to be an idea that is not amenable to physics or computational explanation as a coherent single phenomenon.
Even in a perfectly deterministic universe, the accessibility of that determinism to models hinges on knowledge of the state of the universe.
We're surrounded by highly nonlinear interactions that we just approximate as linear - if we want to truly predict anything, in general, we'd need infinite precision on the variables that describe the state of these systems, because an infinitely small perturbation in the initial conditions can cause an almost (limited by conservation laws) arbitrary deviation in the predicted state.
> assumption that everything about our physical universe is described by computational rules
If you understand "physical" to mean "stuff that physics studies" then the assumption is not so surprising - physics itself is the discovery and application of models to correctly predict or retrodict (i.e. compute) the simple measurable behavior of matter, so it is intrinsically unable to study anything unamenable to computation.
In fact Wolfram is engaged in an attempt (as he sees it) to enlarge the computational paradigm of physics from models consisting largely of differential equations to a "multicomputational" paradigm (see [1]).
I think your statement is much more defensible if you drop the word "physical": The assumption that everything about our universe is described by computational rules is not a given. A classic example might be our inability to sufficiently specify "qualia". If our experiences were fully describable by computational rules I suspect we would by now have discovered how to specify them so that we could agree (or disagree) that 'the red I see' looks like 'the red you see', but our inability to do this is notorious. Nonetheless my (and your) experience of red is part of the universe.
There are physical models are not computable (in the sense of producing number values in a finite algorithm), particularly in quantum physics, e.g. the wave function.
As a physical model, as opposed to a purely mathematical entity, don't you need to add operators to the wave function in order to have applicability to observable phenomena? I'm no physicist, but applying operators to the wave function seems like a computation to me?
I mean, theoretical physics models are still physics models. Functional analysis is still a valid mathematical tool for physics modeling even if it doesn't (usually) produce finitely calculable solutions. Physicists assign physical interpretation to certain parameters, even if they are not (yet) observable.
I suppose there's probably a higher mathematical representation going on here that is finitely computable though. Something like a symbolic representation of the solution functions for the wave function are calculable even if the parameter values are not.
> But that has been challenged by uncertainty principles, observer effect and all that.
Any system which allows universal computation necessarily has subsystems that cannot make perfect predictions about the full system. This doesn’t even require quantum mechanics; humans would run up against this problem in a fully classical universe. Loosely put, since we’re part of the system ourselves, we can’t gather all of the information necessary for perfect prediction without interfering with whatever we’re trying to predict.
However, as far as I’m aware, we haven’t yet closed all the loopholes related to superdeterminism, meaning quantum randomness might not be truly random in the sense that an observer completely outside of our universe could in theory procure an algorithm that predicts the exact eigenvalues for every quantum measurement we make. Since we’re inside the system, no luck for us though. But I like to make a subtle distinction between “unpredictable random” (max Kolmogorov complexity) and “truly random” (ungenerated via any Turing computable algorithm and unpredictable with infinite knowledge). The latter is a bit philosophically difficult to prove exists. The former is quite easy to prove exists; think about how a one time pad has max Kolmogorov complexity (assuming a good source of random noise) but it’s conditional Kolmogorov complexity with the key is much smaller.
Meh? I don't really see this as anything special, or anything close to being an "alien mind". An alien mind would have been trained on its own environment with perhaps wildly different inputs. This is nothing more than a broken human mind. A mind trained on the same inputs as humans, but then altered haphazardly. This teaches us nothing and appears to be purely a gimmick.
He says coherent spots are tiny specs, surrounded by bizarre patterns and eventually vast space of noise and dots. Reminds me of the Universe itself. So much empty space full of fine quantum vacuum noise, and planets and stars… small coherent dots in giant space.
I pay attention to these coincidences, for good or bad.
I always find it interesting how a hero dose of LSD gives similar visuals to what these image AI's do to achieve a coherent image.
on top of that the associated Hallucinogen Persisting Perception Disorder (HPPD) clinical condition seems to be more like an image processing issue, somewhere between the rods and cones (sensor) and memory (persistent storage). like an unwanted configuration change in a program or development environment.
I feel like the more we get AI to act like humans, and the more those engineers and others use LSD, the more convergence we are going to have with curiosity and breakthroughs about how we function. More akin to biological machines without such unique mental faculties, just the illusion or hallucination of such.
This is how our mind perceive or forms reality, ourselves and the objects around us in a huge floating chaos of elements. That's what we call order. Pretty fascinating and overwhelming.
I'd never considered the islands-surrounded-by-fuzz before, but it makes me wonder a few questions - is there a programmatic way to identify the "islands" (without using a separate NN classifier)? Is trimming the "fuzz" a potential method of compression for models?
The dimensionality is a challenge here - these are 2D islands in a, what, 2000D space? Imagine a cross section of a cylinder, that would look like an island if it went through the cylinder, but not along it.
I wonder what it would look like if the cross-section plane was aligned with a line between two recognisable points.
When discussing “n parameters,” such as 6B parameters or in the case of GPT4 around 1.7T is the rumored parameters size, each parameter is a dimension. So, yeah, that’s a lot more than 2d or even 2000d. More like 1,700,000,000,000D.
The dimensionally of these islands must be insane. Every 2D representation is one of thousands upon thousands of possible slices through the weighting space.
Exploring these will be like trying to find the edge of fractals.
Any attempt at visibility into the inner workings of ML models should be welcomed IMO. It's going to be essential in the coming years, if we're going to reason about or regulate them. E.g. how would we hardcode Asimov's laws of robotics into some future deep learning AGI, if it's still just one big black box for us?
Thinking about law (As Asimov's are), I perceive essentially an impenetrable barrier in this possibility space. We create order from the chaos of potential behavior by defining limits.
Currently, in this 2000D StableDiffusion landscape, there are no boundaries for allowable "travel" and you quickly end up in the "sea".
So if we want AI to behave, we should research how to define the perimeter (in thousands / billions of dimensions) and "wall it off" so we can ensure the potentials inside the no-no space are inaccessible.
> There’s also a strange preponderance of torso-only pictures—presumably the result of “fashion shots” in the training data (and, yes, with some rather wild “fashion statements”):
Actually I think this is due to the staggered cropping stage on the input data- the images were all cropped to square, and if an image, like of a person, was taller than wide, the image would get cropped into a few square crops, with one of those being just the center of the image, which, for people in a tall portrait shaped photo standing up, will leave just the torso.
There was a blurb in the SDXL paper about changing the cropping preprocessing to alleviate this potential biasing.
Correct. Very well known artifact, you see it in all the GANs and diffusion models depending on the choice to do center vs random crop. Often changes results by several FID points (random crop adds a lot of variations and is much harder to model, confusing the NNs), which is why I suggested the cropping conditioning idea that SDXL uses with excellent results. (This is another example of how many of OP's observations are essentially minor, contingent aspects of the data or arch, and don't generalize beyond SD1.)
I don't think it's impossible for these images to have some resonance with human dreams and altered states of consciousness. However, the phrase "mental imagery of alien minds" seems to carry unjustified conceptual baggage.
The usual implication of "mind" is something like a process that includes intentions, meta-understandings and etc (varying depending on one's conceptions but usually having these). The processes that generate things could be described as purely algorithmic - or they might have further implications. But the thing there's no proof or even strong indicator of the further implications.
Is this "Ruliad" concept accepted anywhere else? I have a hard time reading Wolfram's expositions, which seem like GEB mysticism (everything is recursion), except here everything is a "Ruliad" (whatever that may be).
Trust Stephen Wolfram to get the idea behind this. I agree completely.
My own work in Stable Diffusion is way less about prompt engineering or approximating 'human art' and much more in exploring its capacity for vivid and plausible hallucination, likewise the beginnings of my experiments with LLAMA.
There is a kind of meaning to these spaces when you can compellingly imply something with purpose and integrity, but it's not prescriptive, or it's not a human thought. I'm looking for the evocative, but I'm willing to be a lot more collaborative about what it evokes… people's explorations into negative prompting are fascinating to me, even though I think it's something of a false path.
Telling AIs to replace human artists and follow instructions while aping their styles is not useful behavior (except economically, and even then it's degenerative). Using AI to do what you'd have to inflict brain damage on an artist to do, is more interesting and ethical :)
I think these tools are much more like a calculator for what is the mechanical part of the human arts - writing, visualization, and others (beyond that I think those use cases are not that interesting, there’s a whole non linear optimization angle I think gets lost in the LLM noise). The “art” they produce on their own with minimal ingenuity is boring. But they also democratize the mechanical skill of rendering images and language, and the intent and manipulation of the semantic space through language, inpainting, LORA, etc - that’s all human induced and human creativity. I think we will find in our future we will be awash in brilliant human arts, assisted by machines, being used to amplify the creativity of people who were before hindered by their skill at the craft of art or the mechanics of writing, much as many brilliant mathematicians and physicists are poor at arithmetic and are greatly assisted by machines in their day to day work.
This is such a fascinating evolution of technology for me. It reminds me a lot of my experience with psychedelics. The ability to see things in a much different way. To visualize the many possible ways an image may truly present to our brains without our learned constraints. Adjusting our model weights. Its beautiful. :)
> I call it the ruliad. Think of it as the entangled limit of everything that is computationally possible: the result of following all possible computational rules in all possible ways.
Quote from his link to ruliad.
I think maybe what he is talking about is equivalent to what we usually call “mathematics”.
Since mathematics actually only instantiates computable things.
(It can represent uncomputable things as abstractions, but only at a level that is actually computable. I.e. everything we know about uncomputable reals, or higher order infinities, is a computable statement. So those concept representations are as computationally rule defined as anything else.)
Once you apply arbitrary rules, you get algebras, so natural numbers, integers, rational numbers, algebraic numbers, continuous fractions, geometry, topology, ...
The view that the physics we experience is a subset of mathematics is obviously not proven, but not original or controversial.
Wolfram has a lot of good ideas. Many of which already existed, but he obscures, renames or ignores that.
He is a genuinely brilliant individual, and if he clearly demarcated what he did that was genuinely original, I think he would both get more respect, and be likelier to not distract himself from doing better work with his own grandiose abstractions.
A lot of great work starts out by solving something small but highly original, and following up every loose end, so that you end up somewhere interesting that neither you or anyone else might have expected. Maybe somewhere profound.
Starting out with grandiose publicly disclosed project definitions creates enormous pressure to confabulate.
Unsure why he needs to adopt the unnecessary axiom of calling existing models "human-aligned". If they were trained from random noise, fresh deep neural nets are the closest things we have to blank slates, which are utterly alien to us and the animal world. Every neural net in nature comes with a bevy of pre-baked patterns that form in the womb and, for humans, continue forming outside of it.
I think the reason Stephen Wolfram is using "human-aligned" is because that is the only feedback loop that is training the models. If you haven't read "What is it like to be a bat?" by Thomas Nagel, I highly recommend it.
Interesting how some of these is very similar to 20th C art movements like Cubism, Impressionism, Pointillism and Constructivism. Picasso et al
Many of which were very much actively practicing altering, expanding or engaging with different aspects or modes of perception.
(edit: also funny how art has its own type of elaborate legalese-type language to (deliberately im-)precisely capture its diverse and expansive meanings)
On the topic of exploring the innards of machine learning models, here are some visualizations I created of the hidden internal representation layers within the StyleGAN3 FFHQ1024 model.
All of your posts are fantastic. I’m amazed you have the patience for the data cleanup required for the plants finetune.
I remember wanting to do a Pokemon finetune (before such things were commonplace) and the dataset cleanup was the worst part. Stability issues and lack of decent multi-GPU setup being a close second.
Thanks! Indeed the cleanup required for the 'Nature Artificial' dataset was intense and forced me to consider other techniques. It was around that time that I realized that I could use Stable Diffusion to output 10,000 to 50,000 images, which has proved to be a useful dataset creation tool for training StyleGAN2 or StyleGAN3. Yet nailing down a SD text prompt that will continuously output images in a specific style and context can be challenging.
StyleGAN2/3 is quite sensitive to fine-tuning models with a dataset that contains lots of variation, such as the Pokemon dataset. But it will converge much better with a dataset of just Pikachu images in tons of different poses. It seemingly wants to find the most common pattern in the dataset and interpolate within that space. From there nailing down the ideal value for the gamma attribute is difficult, but luckily can be reduced every 1000 to 2000kimg until the model no longer trains favorably. Currently my typical gamma strategy is: 80 > 10 > 5 > 2
Our models have the be the furthest thing from an alien mind. Almost mathematically. A trained model with no hallucination is essentially an average of all training data. Which comes from us. Who are not aliens.
Can see Wolfram's approach here - interesting how he's coming to generative AI with a beginner's mind, from first principles. He makes a lot of interesting points, but I can't help and think that he would save a lot of time talking to someone who knows the field better, instead of figuring it all out on his own.
The stuff about attractors and dimensions in model space is missing the architectural understanding (layers representing features, backprop)
That's weird. A cat is a cat is a cat, any alien with a visual organ and a perfect ability to paint, would reproduce the same thing.
Inside his mind he might have completely different associations with the image, and this weight manipulation trickery doesn't get us closer to those hypothetical associations.
An infrared image of a cat an an X-ray image of a cat are drastically different to an optical one, so it's not a given that they'd all draw the same thing.
If the alien is blasting the cat with soft x-rays constantly to perceive it, then it's mostly different because it killed the cat with radiation poisoning.
Shapes and proportions remain the same regardless. And even human artists can be... "inventive", with coloration or emphasis of some features over others.
It seems improbable that they'd draw anything we haven't drawn ourselves.
If you asked me to draw a cat in a party hat, I would get the number if ears legs and eyes correct, but my fur would be non-existent and there’s no way my cat eyes would glow in a reflection. The AI is thinking differently than humans.
Then the AI got better, and still the dreamlike things never go away. The constant problem with the hands, words never properly formed, clocks always look wrong, etc. But now it is more cohesive, more "solid" but still malleable. Like a lucid dream. Still straddling the border between consciousness and unconsciousness but there is a hint of control and direction now.
As someone who lucid dream, I think the AI images and the LLMs are just like a person dreaming right now. They sort of have control, but actually not really. Kinda hard to explain but even when I lucid dream, I know it is a dream and I can bend it to my will but at the same time, it isn't possible to control my thoughts. Trying too hard to assert control and I wake up. So it is still a state of unconsciousness for me and not at all comparable to the "me" when I am fully awake.
Of course, the AIs can't wake up if we use that analogy. They are not capable of anything more than this state right now. But to me, lucid dreaming is already a step above the total unconsciousness of just dreaming, or just nothing at all. And wakefulness always follows shortly after I lucid dream.