Are AI models narrowly creative?
Showing only #4683 and its comments.
See full discussionLog in or sign up to participate in this discussion.
With an account, you can revise, criticize, and comment on ideas.
AIs have created output that is not only novel, but seems to constitute new knowledge (resilient information), such as the famous Move 37 from AlphaGo. That is new knowledge because the move was not present in the training data explicitly, nor did the designers construct it.
Move 37 was not new knowledge. It was the winning choice in that situation before the AI ever existed, because it was deducible from the game's rules and the current board state. It was implicit knowledge, already contained in the system at that time. AlphaGo made it explicit, by finding it, like a search engine, but did not create it. If you calculate the trillionth digit of pi, you haven't created new knowledge, at least not in any sense we should mean. You have simply revealed a value that was already fixed by a definition.
The fact that Move 37 wasn't explicitly in the training data or the programmers is irrelevant to its status as knowledge. This is true for pi, and for all content created by AI at the time of this writing.
If there had been no AlphaGo and no Move 37, and a human had made that move, as they have similar moves, it would no doubt be called creative genius (as similar moves have). Isn't the above a double standard?
If the human made Move 37 for the same reason as AlphaGo, it would not be creative. Such moves are creative when humans make them because they are not deducing them (they can't due to practical limitations). If something can be deduced, it is not creative. Creativity is the conjecture of a new structure which is not derivable/deducible/implicit via existing rules of inference. All AI-generated art is implicit in the training data and model design in the same sense, so is not being made via creativity.
This highlights the core mystery of AGI/creativity: if it is the creation of something which cannot be deduced from existing rules (yet is still helpful, hard-to-vary, knowledge-bearing, etc.), how can it be programmed? In a sense it cannot, as Deutsch writes: "...what distinguishes human brains from all other physical systems is qualitatively different from all other functionalities, and cannot be specified in the way that all other attributes of computer programs can be. It cannot be programmed by any of the techniques that suffice for writing any other type of program." [https://aeon.co/essays/how-close-are-we-to-creating-artificial-intelligence]
The definition of fitness that rendered Move 37 the best choice originated outside the system.
The definition of fitness for DNA also originated outside it, so this doesn't in itself suggest the system isn't actually creating new knowledge.
This seems to me to be the same distinction that Deutsch and others have made between the genetic evolution we can simulate through evolutionary algorithms and the kind we actually observe in nature. I think it would be helpful to investigate evolutionary algorithms a bit further if you want to develop a clear distinction. This is how I describe it in my book:
There are several mechanisms that genes use to create variants, including sex, mutation, gene flow, and genetic drift, all of which appear to introduce change randomly. But we now know it cannot be entirely random. Something more is shaping what gets trialed, because when we model and simulate evolution using random changes, we never see the sort of novelties that arose in nature. We see optimization. We see exploitation. We see organisms become better at using resources they already use. But we never see a genuinely new use of a resource emerge. A fin may become better at swimming, but it does not become a limb. A metabolism may become more efficient, but it does not open up an entirely new biological pathway. And yet the natural world is full of exactly such extraordinary adaptations.
I keep returning to the notion of the space or domain in which simulated evolution so far operates in. It seems like we can say that current sim'd evolution can discover new knowledge via conjecture and criticism, but it is always bound by a domain predefined by fitness functions, automatic evaluators and so on, even if that domain itself contains many subdomains.
Then we can say that in nature, and in the minds of people, there is no externally defined space in which exploration is happening; the space is also evolving, also subject to criticism. I suspect this is part of how open-endedness comes about.
But the immediate question here was how to explain why AI is or is not "creative". Saying AIs are "narrowly creative" seems it could work, or saying they are creative within a fixed domain. The common intuition I think is that current AIs are "truly" creative, and I would say this is because the predefined domain (of LLMs, for instance) is gigantic, being sculpted by an internet-sized training corpus. But I suppose we should argue that "true creativity" means universal creativity.
I was curious if there are criticisms of the argument that current AI does legitimately create new knowledge.
Move 37 was not explicitly present in the training data, nor designed by the programmers, and is extremely hard to vary (Deutsch's criterion for good explanations). Was the move present implicitly in the design of the system and/or the training data? Or inexplicitly? Do either of these mean the discovery of the move was non-creative?
Since evolution created genetic knowledge from nothing, it can be said to have the same "narrow creativity" as AI. The confusion over whether AI "is creative" can be resolved by saying that it is, but only narrowly (like evolution), and that the creativity defining people is universal, not limited to any domain. AI creates knowledge in domains it was designed for; AGI can create knowledge in all possible domains, each of which it designs itself.
Criticized per #4718: AIs are not "narrowly creative"; there is only creativity in the binary, universal sense, per Deutsch.
This also admits of the distinction between AI and AGI (and "universal creativity") as being whether the system is capable of creating knowledge ex nihilo, as argued by Deutsch. Only universal creativity could create knowledge from nothing. Bounded creativity must start with something.
But nature created genetic knowledge from nothing. So this is an example of something which does not have universal creativity which created knowledge ex nihilo.
Nature does have universal creativity; it can generate any possible knowledge. And all possible knowledge exists somewhere in reality.
By this standard, a random number generator has universal creativity as well, and is therefore a person. So there must be a standard for personhood other than: able to generate any possible explanation. Such as: can do that tractably.
I think tractibility lacks the open-ended capacity to reformulate what counts as a problem, a solution, and relevant data. Creativity is (at least partially) the ability to reformulate the problem space itself, not by ironing out implications of existing theories. An AI and computational systems is already good at ironing out the implications in our language and existing knowledge systems. But that's search within a given space, not the creation of a new one. Creativity seems to work on a higher level. It's operating at the level of problem framing, which requires things like relevance. An AI can't create new relevance, because its weights are a statistical compression of what humans have already found relevant. It inherits a pre-given frame.
I might be confused about what you mean by tractible. But it seems to me that tractability can't do the work the bounty asks. Tractability is formally defined relative to a fixed problem space. But universal creativity is (at least partially) the capacity to restructure the space, to change what counts as a problem, a solution, and relevant data.
An interesting example from cognitive science is the Mutilated Chessboard Problem, which asks whether a board with two same-coloured corners removed can be tiled by dominoes. As a tiling problem the search space is combinatorially explosive. But reframe it as a colour problem and the answer is easy. Every domino covers one black and one white square, and you have unequal numbers of each. The solution came not from searching harder, but from seeing the problem differently.
I agree that tractability is related to a given problem space, and that creativity is about reshaping the problem space, among other things. Given that I've been thinking of the problem space as the space of all explanations, I'm not sure where I stand... Maybe the "space of all explanations" framing is wrongheaded, because a mind never has any actionable knowledge of that space? We can discuss the space of all explanations in some sense, but we can't organize or describe it in any substantive way...
Also, per #4865, you helped me remember that personhood could involve intractable algorithms, but ones which only ever run with small inputs, since that can still be perfectly practical. Whether or not that means the whole person is a tractable algorithm or not, I'm not sure.
Between these points I think this is enough for you to claim the bounty, because it does argue that personhood "should not be defined in terms of tractability", per the bounty terms (italics mine, here). Tractability does not help explain personhood. Or, in any case, it doesn't seem like this line of discussion will be very fruitful (but this could itself be mistaken).
Upon review, we should maybe say instead that personhood should not be defined solely in terms of tractability, which the bounty terms are not clear about. As it stands (bounty aside), I find myself still seeing tractability as an important aspect of epistemology and the mystery of personhood/knowledge creation, a hunch reinforced as I continue reading through "Why Philosophers Should Care About Computational Complexity" by Scott Aaronson: https://www.scottaaronson.com/papers/philos.pdf
Knut, you’ve won the bounty. You need to integrate with Stripe to get paid.
An alternative criterion for personhood is speed: a person is a program that can synthesize any explanation in less than the lifetime of the universe, say.
This is a bad criterion because then random program generators are sometimes people.
This wrongly implies speed is a property of programs, but it's a property of hardware.
Speed is a property of programs, too. https://en.wikipedia.org/wiki/Big_O_notation
Ah, so if I understand correctly, there are two knobs affecting speed (elapsed time) for a given algorithm: the hardware, and the implementation of the algorithm. The given algorithm has a complexity, independent of those two, which is how the time and memory scales with an input.
The given algorithm has a complexity, independent of [the implementation]
No, the complexity depends on the implementation.
"Complexity" in the sense of growth behavior with input size? Further reading is still suggesting to me that this is intrinsic to a given algorithm (or class of them). Intrinsic to the math and logic. Implementations can be faster/slower/hungrier for a given input, but if they have different limiting behavior, aren't they different algorithms? I can see how an "implementation" of one algorithm in practice can accidentally change it to another algorithm.
"Complexity" in the sense of growth behavior with input size?
Yes.
I can see how an "implementation" of one algorithm in practice can accidentally change it to another algorithm.
Not sure why you put that in scare quotes. You might be right in the CS sense where ‘algorithm’ refers to an abstract procedure whereas ‘implementation’ is concrete code realizing that algorithm. (Though as a disclaimer, I don’t have a CS degree. My experience with programming is fully on-the-job.)
My point is more that two different implementations that compute the same function can have different big O. In that case, they’re usually considered different algorithms, even if the high-level goal is the same.
Regardless, the structure of the program is by far the most important factor determining performance characteristics. If you were saying that complexity is independent of implementation only insofar as the implementation truly implements the same algorithm, then I agree. So I’m not sure whether I should mark this as a counter-criticism. For now I won’t, pending new evidence.
I think I see now, and agree with the above. Partly a semantics issue (yes, I'm thinking of an algorithm in the "formal" CS sense: an abstract/mathematical finite procedure). The scare quotes were meant to suggest that one could attempt to implement one algorithm, but the implementation may in fact be more closely implementing some other unrelated algorithm, but this is confusing.
At any rate, how ChatGPT summarized it makes sense to me:
"One function → many algorithms can compute it.
One algorithm → many implementations can realize it.
Complexity attaches primarily to algorithms, secondarily to implementations, and not to functions."
"Secondarily" meaning:
Implementations of an algorithm inherit the algorithm’s asymptotic behavior. If an implementation has a different asymptotic behavior than one algorithm, it is effectively a different algorithm.
A random number generator does not have universal creativity, because it is not a universal explainer: it can only generate explanations by accident. Universal explainers seek good explanations through conjecture and criticism.
Understanding explanatory knowledge seems like a better criterion
Maybe... but "understanding" is too vague, I think. Doesn't understanding mean: can explain? But then this is just "can create any explanation" again. I think the core question is why a random program generator isn't a person, coming from Deutsch's definition of a person as a program that has explanatory universality -- can create any explanation (my thought here is that this definition isn't good enough on its own, given the random generator point).
"Understanding" isn't just another way of saying "can explain.". Explaining follows from understanding, but isn't synonymous. An RNG could by chance generate a good explanation, but it doesn't understand it, and therefore can't distinguish it from garbage. Understanding involves recognizing that something is a good explanation. It is conscious understanding that makes conjecture and criticism possible. Without it, you have no criticism, only random selection. What do you think of the suggestion that what's lacking from the explanatory universality definition, is an intelligent selection mechanism. A random program can generate any explanation given infinite time, but it will never select which explanation is good.
This is a good point, related to Dirk's #4813. As far as the bounty goes, I think my response in #4823 applies here as well, however. To refine it:
Recognizing, criticizing, and being able to understand explanations could all be requisites for tractably synthesizing any possible explanation. The bounty regards whether the tractability requirement can be done without.
It seems like a mind being able to create, recognize, understand and differentiate (etc.) good explanations are necessary but not sufficient criteria for personhood; if that process is intractable, then beyond a certain amount of current knowledge (considering that as the input to the process), the person effectively cannot continue with it... so that compromises the universality.
They must be able to create, recognize and understand any given explanation, and maintain that ability as their knowledge grows, ad infinitum...
By Tractible, do you mean "efficient relative to fixed task"?
Yes, my understanding is that the standard sense of tractable, for some algorithm, is: can be executed in time that grows at worst by a polynomial function of the input size. This is the sense I mean. The fixed task would be: create a given explanation in the space of all possible explanations.
Implementations of a given algorithm can be way more or less efficient in practice, though. Maybe personhood does require intractable algorithms, but ones which only ever run with small inputs... The question of the bounty is whether can we make a case for or against this. But part of the hope is also to learn if this whole framing is mistaken.
Doesn't it? All explanatory knowledge is in the set of all possible programs, and a random program (or number) generator can generate any of those, given infinite time.
We could say a person is a program that can synthesize any possible explanation in finite time, excluding memory limitations. But this would again grant personhood to RNGs. For that matter, a counting program could just enumerate all possible binary strings up to its memory limit, in finite time...
Creativity isn't defined by its outputs but by its process. RNGs do not recognise or criticise ideas.
Agreed on both counts, but I think the bountied idea survives this...
Recognizing and criticizing ideas could be a requisite for tractably synthesizing any possible explanation (I suspect as much).
Tractability is a consequence of creativity. It's a little like saying the difference between you and a rock, is that you can move faster.
Universal explainers seek good explanations…
You sounded persuaded by https://blog.dennishackethal.com/posts/hard-to-vary-or-hardly-usable. As in, you agreed that people don’t seek good/hard-to-vary explanations.
So why still speak of good explanations?
Universal explainers
In the context of how AGI may work – which seems to be what Tyler is mostly interested in – the concept of a universal explainer might not get us very far. Creativity is the more fundamental concept, I think.
A person is a universal explainer, yes, but he could also use his creativity to come up with reasons not to create explanations.
By the latter standard, neither nature nor random number generators are people, which is sensible; nor can nature create any given possible knowledge tractably -- this is true because the fact that all possible knowledge exists is only by way of the multiverse, which is a process that cannot be simulated in its entirety, even by a quantum computer.
This also admits of the distinction between AI and AGI (and "universal creativity") as being whether the system is capable of creating knowledge ex nihilo, as argued by Deutsch. Only universal creativity could create knowledge from nothing. Bounded creativity must start with something.
Moved the criticism of 4694
Implicit vs Explicit knowledge -- AI, like DNA, has the ability to create implicit knowledge.
Not to be a stickler but I think you mean ‘inexplicit’.
Implicit = not said directly but implied. Can still accompany explicit speech though.
Inexplicit = not expressed in words or symbols.
At least that’s how I use the terms.
Cool - maybe that is better. I think that inexplicit knowledge is underrated in the David Duetsch circles, what do you think? ... there, I used it in a sentence, now I'll remember it!
Nice, yes. I do see Deutschians using the concept, especially in the context of the fun criterion. But in the general public inexplicit knowledge is underrated, I agree.