Regression to the Mean: on LLMs and the quiet death of the new(rruxandra.github.io) |
Regression to the Mean: on LLMs and the quiet death of the new(rruxandra.github.io) |
The problem is that this is a contradiction, and a pretty common misunderstanding. When we talk about something being probable, we're talking about what we don't know. When you extrapolate, you're saying something about the unknown, you're creating something new. While you can extrapolate into the past or the future, in the way this line is talking, it's answering in a future tense. I think even worse is that beneath this it says
> returned at the speed of certainty
At that point we're no longer talking about probabilities!
The problem isn't that these machines aren't capable of making new things. The whole of their mathematical grounding is in the creation of the unknown from the known. The problem is precisely that they are sold as miracle cures where they can produce great results for little effort. The law of "you get out of it what you put into it" still holds true. Undirected, uninspired usage of these statistical models gets you mediocre at-best results. Without an understanding of the underlying theory and mechanisms of how these models work (it's not just transformers, but any statistical model), driving the whole of the inference chain with maximal control as one might Max/MSP, as well as mastery over the target domain, you will effectively achieve nothing but "slop".
Of course, there's a whole other discussion here, which is that this site seems to be victim to the same grave ignorance that has caused the supposed "crisis of newness" within the arts (which has been talked about for much longer than these models have existed). That's a whole other can of worms, but essentially it's bunk. In modernity we can point to the last century of unending artistic innovation, and panic that this is slowing down, that this is the end of history. In truth, that century is anomalous. It's the most anomalous we've ever recorded, where real material changes were reacted to in real time. The innovations of modernism weren't born because of pretense to being original. It was wholly derived from the changes happening in reality irrespective of the arts, as a result of the industrial revolution, and later the information revolution. The norm in history is centuries of very slow refinement, barely perceptible on the timeline of a generation. Tiny little incremental changes stacked up over a long period of time. Bombastic, revolutionary artistic progress is the anomaly. An unending cacophony of that progress has happened exactly once in the entire history of humanity, as far as we can tell. There is a stupid expectation that the 20th century's breakneck pace was going to last forever. Obviously it wasn't. It was never a sustainable momentum, statistical models or not. People are still in the mindset it's the norm. The languishing over creative bankruptcy is simply the death of this delusional fantasy.
(Or am I now like the Dwarfs in The Last Battle, seeing all things as mere simulacra?)
> A running transmission of half-formed thoughts, sketched before they cool into language. Pages that perform their ideas instead of explaining them.
To take a simple example: I grew up with computer games in the '80s where there were no 'physics engines' or frameworks for building games. As a result, each game was an expression of the author's personality somehow. Fast forward to the noughties, games bored me as they mostly looked and felt the same, or maybe felt like 3-5 different games all packaged differently.
Another example: going abroad on holiday in Europe (I'm from London) used to be a relatively wild, vibrant experience, filled with unexpected differences and challenges (not all positive). There were no McDonalds or Starbucks and the shops were filled with unfamiliar products and foods. Now everywhere in Europe feels the same when I visit, especially with smartphone in hand.
And films went from wildly different to one another to what now feels like 'arty' and 'CGI' being the two choices.
This article continues that into the realm of ideas, or idea production. Everywhere you go looks and feels familiar.
Or am I just getting old?
Perhaps? But I think this is more a case of just not seeking things out.
Music is as vibrant and diverse as ever, but not if you're only looking at the top charts run by the music industry.
Same deal with games, there's more experimentation and interesting concepts in gaming than ever before, but not from the AAA studios.
Now I can't speak for how you vacation, but I've had wonderfully different experiences between Hamburg, Berlin, Vienna, Prague, Rome, Paris, Montpellier, London, Amsterdam, Oslo, and Florence. I just don't go to the starbucks and instead wander around a bit, optionally picking from a few hit destinations if I feel like it. But also, it's not like this was created for nothing: https://www.itchyfeetcomic.com/2018/10/omnimappus-europeus.h...
I just wanted to add on to this, I wouldn't really classify modern games as "lower quality" than those of the 80s. I'm really not a fan of AAA games, I think the last one I played was Elden Ring, but I would never suggest that they're actively low quality. Uncompelling? Absolutely. But I also have spent a lot of time playing games from the 80s. Silver and golden age CRPGs, random simulators, DOS games that catch my eye. "Quality" isn't the first thing that jumps to my mind. Often they're ugly, terribly balanced, buggy, rife with all sorts of issues in any category you can think of. Games have come a very, very long way. 2400 AD (1988) and Champions of Krynn (1990) are relatively speaking highly polished masterpieces. They're still kusoge, honestly. I have very little experience with the consoles of that era, because pretty much nothing I see even remotely catches my eye.
I don't think people under a certain age can really appreciate how different it was.
Both games and movies are predictable in the sense that we know what to expect, and they have been largely standardized. Games have common keybinding schemes, as well as user experience mechanics: how jumping feels, when we expect to autosave, what the UI/minimap symbols mean, etc. When it comes to movies, I find myself no longer turning away from the screen before gruesome scenes, because I expect in advance that they won't show it, depending on the mood of the movie. I also find that you can often predict which dialogue lines were meant as foreshadowing for a plot twist coming later. This standardization is intentional in the sense that people are more likely to consume something they are familiar with, and more likely to enjoy it if they can passively engage with it.
It's common nowadays to pay $20 for a game, play it for a few hours, and forget about it. Or, turn on a random Netflix show on the TV to pass time in the evening. Quite likely that a month later you won't reminisce about either of these experiences, but you probably didn't have high expectations either way. I think 'consuming' a travel trip is similar in the sense that it has very familiar tropes no matter where you go, but more implicitly resulting from market forces rather than intentional design from a creator.
The Binding of Isaac: Rebirth, Bit. Trip, Cave Story, Crypt of the Necrodancer, Cuphead, Downwell, Fez, Hollow Knight, Limbo, Octodad: Dadliest Catch, Papers, Please, Proteus, Risk of Rain, Rogue Legacy, Shovel Knight, Spelunky, Stardew Valley, Super Meat Boy, Terraria, VA-11 Hall-A, VVVVVV
Can you tell I own a PS Vita?
There's going to be reams of AI slop (already is), but I bet the amount of amazing games will also (more slowly) increase due to AI tools. The trick is in how well we can filter.
I think we're in the early stages and being overwhelmed by low quality production. We'll find ways to filter, and find some real bangers.
It's true that in a project, a novel idea undeclared as such will be shaved off quietly by an llm. You really need to be explicit about wanting to keep it.
You will get pushed into the mean.
However, I'd say 90% of making something (that is useful) is repeating the old thing. We stand on the shoulders of giants. Or at least we should. Getting there can be difficult for most of us.
I say this as someone who chronically re-invents things. I then later get stuck and find someone already thought through my problem and solved it better.
I don't believe being unique in all the ways is useful. You need to be unique in the important ways and not unique in the other places.
There's also a cultural coherence angle that (my) unique things often fail at. Stuff has to look like other stuff enough for people to understand intuitively what it is and how it works. Here the mean is your friend.
I am able to explore more unique spaces because I no longer deal with the minutia of getting the things that should be the same correct. So paradoxically, this has made my output more unique.
The entire point of software engineering is to make repetition unnecessary so people can focus on the new.
LLMs are pushing the people that use them into the worst possible set of practices.
Now however, if people are not even internalizing those fundamentals in order to even re-write things in their own words, using their own "mental model" (perhaps correct, sometimes not) - I fear they won't even develop "mental models" and abstractions...
There's a very interesting opportunity here for a deeper investigation.
- What does "regression to the mean" actually mean in practice when the LLM is conditioned on a possibly large amount of context?
- How does this perceived regression to the mean affect the result in different applications? When implementing code, it may show up as keeping it simple, hence easily understandable, "nonclever". When writing documentation, it may show up as simple language, short sentences, etc. supporting the intent of communicating with little friction to a broad audience. When brainstorming product ideas, it may show up as regurgitating old and boring ideas, but dressed in fancy language and affirmations that hide the shallowness of the content.
- What can be done to alter this behavior? Now that temperature doesn't seem to be a parameter anymore in new models, how can we steer creativity of the model?
- If the model's creativity is fundamentally limited, is there a way we can use it to support us in the expression of our creativity, leveraging the different strengths of humans and LLMs in a way that the result transcends the limits of either?
Unfortunately, I don't see the article doing that. And, while I know pointing out LLM-isms is often a cheap shot these days, I feel compelled to point out that this article is full of what I perceived as LLM-ism, quite ironic given the premise and the statement ("written off-distribution · on purpose").
E.g.
> Trained on the past, it answers in the past tense of thought. Not what is true. What is typical.
> We converge — not on what is right, but on what is average.
> Not the answer it was sure of — the one it would not stop correcting
For the purpose of discussion though, this also undersells AIs:
1. They CAN be great tools for novelty + discovery! You just need to ask and explore and put in work. Its not "easy", but it does help.
2. Sometimes "the mean" is what you want. Sometimes I'm not after art. I'm after something efficient and recognizable and easy to maintain.
How this oversells AIs:
1. We are losing the muscle of forced creativity and problem solving. There is a certain kind of learned privilege that comes from facing a problem and having your instinctive reaction be to ask for and expect help from something else rather than to roll up your sleeves or sit back and have a think. If the system incentivizes loss of muscle en-masse, we're gonna lose something beautiful and powerful.
Many folks have touted the "calculator" similarities as an argument, saying it's more of an efficiency gain / productivity enhancer. To me, LLMs are far more involved than this. Now, unknowingly (or knowingly), people are offloading the problem solving portion of small tasks.
- Creative Writing (Claude, make this email sound more professional)
- Coding (Handle this small logic bug for me)
- Note Taking (Generate a summary of this meeting recording)
- Strategy (Set up a roadmap for X project) and many other areas
- Design (Give me a powerpoint for a stakeholder meeting)
- Personal Life (Find a restaurant I can take my wife to for our anniversary)
Many people underestimate how many "simple" tasks required creative problem solving abilities, and we're actively handing more and more of that over to the thinking machine.
Perhaps it's human nature to give this up, and maybe it's in our best interests - but this is the first time I've ever seen people stop thinking for themselves en masse. Interesting times ahead, IMO.
99% of the time we don't need a true intellectual breakthrough to get the job done, and often 'new ideas' are simply riffs on or blends of old ones, like fashion or music genres.
The worry to me, however, is that if society comes to rely on this form of 'AI' then eventually the model collapse bleeds into academia (e.g. grant proposals reviewed by AI?) causing a kind of incremental sociocognitive atrophy. Everything becomes a reaffirmation of the status quo.
That being said I think people said something similar about electronic calculators (that if you couldn't do long division by hand then you'd be too incompetent for higher-level calculus.)
https://mapwriting.substack.com/p/living-subscription-free-i...
It's the same problem that AI faces of Model Collapse: AIs that train on the internet ultimately just end up training on one another, stop moving forward, and end up as identical polished versions of one another
I now think of it as a Dr. Jekyll/ Mr. Hyde situation for software projects:
- Dr. Jekyll: For makers, the only limit is your imagination, architectural guidance, and token budget. Time to build!
- Mr. Hyde: For projects to get off the treadmill of having to copy others to maintain you position, you need to redefine how the project works and provides unique value. Features and quality are no longer the answer. Time to fight!
Similar to a human that wants to learn something and invents exercises to practice.
1-2 years ago it was a theory, but new models are trained, successfully, on synthetic datasets.
- Synthetic datasets are typically human-steered today, which points to model collapse wrt learning from the internet. I don't think standard practice is (yet) AI looking at the internet and deciding to build its own gyms to go further. When it does, model collapse may happen again, and be even more expensive
- Distillation attacks are getting interesting here. There seem to be 2 kinds: intentionally querying other models, and maybe not so intentionally, learning from reasoning traces going through shared routers, esp. coding ones
- A lot of neolabs are trying to go where the big labs might not look as directly to avoid being squashed, which suggests they aren't ready to bet on being smarter, and that means the general AI is more about collapse / $
But, there is some truth to the article and perhaps it is more true in chat based interactions. The agentic, hands-off mode might tell a different story.
I disagree. LLMs are essentially trained on all the math we know, as well as - via RL training - how we apply it to solve given types of problems.
It should be obvious that not every known mathematical technique, or sequence of techniques, have yet been tried by humans to solve every unproven conjecture, so necessarily there is some amount of low hanging fruit where an automaton tasked with throwing the kitchen sink at a problem will be successful. This is not the same as creativity.
If you are familiar with the mathematical notion of a closure, then what LLMs are capable of is generating the generative closure of what they were trained on. They can generate output that is novel in the sense of a sequence of chess moves, or mathematical moves, that was not - as a sequence - in the training set, but yet it WAS in the training set as part of the implicit closure of what an LLM would be capable of generating.
But yes, LLMs are likely to force permanent conformity.
One could also talk about how language in general shifts with the population, but LLMs are likely to prevent it. One would think anthropologists are already looking into this experimentally...
If you bring something new to the table, then in my experience, AIs are really good at helping you ground it old ideas. If you want to set it and forget it, then you will get the mean. If you want to do something new, in my experience, they are enablers and not blockers.
I think that perhaps there's a bit of hope, that by the forces of the market, the value of human distinctiveness will rise in comparison to whatever is the generated mean. This is what I am looking into.
The pushback is soft, and constant: Did you mean: the familiar thing, offered in place of yours."
Ouch, that's scary to think of.
Also why does it read like its written by ChatGPT?
First of all, neural nets do nit return averages per se. They construct space between the points and extrapolate outside of the points. So even if a point was not in their training data, they will be ok, in many situations, to acknowledge it.
Or in other words - LLMs don’t average. They construct world models. A novel thing that fits their world model will be accepted no prob. A thing that doesn’t may still be accepted but with challenges.
The same is true though for humans, including scientists. There is a saying that science moves one grave at a time - because often prev gen of scientists needs to die off for a new idea to take root.
Or in yet other words - even if llms produced averages, an average of a discontinuous set can lie outside of that set. And the set of all human ideas is very much discontinuous.
They don't. They interpolate between the points on a manifold.
The argument stands.
is the proof to the Riehmann hypothesis also somewhere on the manifold and we just need to prod the LLM with the right prompt so it locates the point?
Anyway, going to read the actual piece but felt I needed this off my chest.
The simple fact of the matter is that these things are best viewed as a computer implementing a continuous form of computation. If you can understand the following words: "the function between function 1 and function 2 is function 1.5" and can imagine that this is a process with "infinite" descent, where you can continue to pull functions halfway between other functions trivially, that's a pretty great mental model to have. To that end, using them is like operating a big, complicated radio setup. Or a huge collection of synthesizers and filters. You're essentially tuning in to a structure. It can't help you figure out what structure you should reach for, or what you should do with that structure.
The creativity doesn't come from the tool itself. If you were only capable of having a mediocre body of work before, they aren't going to help you after. There is no removing human brilliance from the equation. AI is immensely exciting for what it promises, there is something genuinely new and interesting here, but it's not a crutch for the untalented. That's how it's getting sold, and it's having a massive disservice done to it. What we have is a genuinely new space to pioneer in. The interesting stuff and compelling art isn't going to be found at the end of a however-many-word positive prompt, but a symphony of total control over the model itself. Drawing 50 original works for the purpose of fine-tuning a diffusion model for a single project, and only as one particular component of the fully customized inference workflow, which will be scrapped when the project is completed. That's what artistic usage of statistical models looks like. The construction of a specific program for a specific purpose. Anything else is fantasy. The model can't give you that purpose.
The point w/ electronic calculators is the same point made by Plato regarding books. It used to be easy to laugh off these concerns, not so much today. Imo, this is the real progress: people are now asking meaty questions regarding the ultimate human purpose of books, calculators and technology.
But the people studying math and the related fields are able to do division by hand on paper. They are just slow when doing it.
I believe that the calculator was meant to solve the slowness problem rather than eliminate the need to fundamentally understand division.
That's a common meme but it's the opposite of true. Everything big models, not just transformers, mathematically do is extrapolation in the feature space, almost never interpolation. They're perfectly able of combining the ideas, although of course this ability declines once they're off the distribution, just like in humans. The model is creative and its output is transformative, however it only makes sense if you define creativity in a pure manner, based on novelty for itself.
However most people use entirely different definitions of creativity, something like "surprise me in a way that still makes sense to me". This includes "me", a side observer, and depends on what the side observer considers novel. Hence the main reason for the lack of "creativity" in big models is not some hand-waved "regression to the mean" or "interpolation", but the fact that they're still insufficiently intelligent compared to a human. Current models simply don't have enough fidelity to understand humans and think as deep, that's why humans think their output isn't sufficiently novel for them! Think about this, insanely smart aliens would probably dismiss humans as non-creative too.
The contributing factor is also the lack of semantic diversity in current models. Also known as mode collapse, but the name is a bit of misnomer and describes a technicality, not the resulting phenomenon. This indirectly affects creativity as it's usually understood, because the models generate and repeat the same thing in response to the same thing, which is the opposite of novel in layman's definition. That's part of where the slop comes from. Mode collapse has many causes, e.g. post-training with current algorithms. It's likely fixable but AI shops show little to no interest in studying and fixing it.
But you're restating what I just wrote - We're training a status quo machine and the probability of anything outside that distribution rapidly drops to zero.
I've been playing a survival crafting game lately called Abiotic Factor and it is blowing me away with how creative and fresh it is, even in the fairly crowded "survival craft" genre
Cities could look like https://en.wikipedia.org/wiki/Habitat_67
Cities could look like https://www.atlasobscura.com/places/the-blue-city-of-jodhpur...
By design this must be the case, even when you account for stochastic sampling i.e. 'temperature'. All it's outputs are a highly-dimensional combinatorial interpolation (I'm talking about GPTs here)
That's probably why Claude is very good at producing plausible nonsense rather than the often correct response of 'I don't know'.