Merged Artificial Intelligence

[Chanakya]

You have been answered clearly

No, I really hadn't been, not so far! But it's cool, no issues. You've done that now, with this post, as far as the first of the two questions I'd been trying to explore.

- the current AIs can produce content that has not existed before in this universe (or at least in our light cone). The question you seem to want to ask is really are they “creative” in the same way humans are “creative”?

No no, what I was asking wasn't so very nuanced. In fact, I don't even know what that nuance actually means! Your first sentence answers my question fully. That's exactly what I was asking, whether that story of yours, for instance, is the derivative "slop" that's everywhere these days, or if it is actually original, as in not wholly derivative.

But I'd just like to explore your view of this just a bit more, if I may. I mean your view that AI can indeed produce original work. Just how original, that part let's explore just a bit.

Where I was coming from is this: There's some/many humans that produce work that's technically original, but for all practical purposes derivative. That's human slop. (You see what I'm saying? If Roko had basically just pieced together bits and pieces, not just something only somewhat related like Pascal but actually closely similar: then, while that wouldn't have been original in the sense I meant, but it would still have been original human work if you take the whole bunch of ten, or twenty, or fifty humans whose work Roko drew from. ...Can AI do that, is what I was wondering: can just one AI, or even a whole bunch of as many AI bots as you like, get together and put something together that's original, in the sense of not being derivative, in the sense of not simply being an amalgam of what it's seen a lot of humans right.)

And you answer that with a definite Yes now. Which is cool. I'm happy to take that onboard, absolutely. ...Well, provided that: (1) We're talking about the same thing, see my paragraph above; and (2) If you're sure you're sure about this, if you know what I mean.

And the answer to that is a clear no, current AIs do not think like humans, they mimic human behaviour with different processes than what produces our human behaviour. But why do they have to be the same as humans to be “creative”? You’ll note I’ve used scare-quotes around creative, that’s because 1) it is a very ill defined word 2) I think the current generative AIs have revealed what we used to consider the domain of humans, that required something special (as in biological process not a soul or any form of dualism) is not the case.

Eh, this, no, this isn't what I was referring to at all. Where I was going with this was more straightforward, like I explained earlier in this post. This, what you're talking about here, I'm afraid I don't even know what that means, really! ...But still, although wholly different than what I was talking about myself, but still, this is interesting, I'd never given this any thought! ...Maybe you could spell this out a bit more: What would be a story-or-scenario-with-human-thoughts be exactly, and how would it differ from a story-or-scenario-with-non-human-AI-thoughts?

(My definition of “creativity” is producing something coherent that has not existed before. I’ve posted several examples of this, including the program I created entirely with much older - 6 months or so ago - versions of AI, that program had never existed before.)

Mine as well. Except, I'm just further trying to distinguish merely technically original but entirely derivative --- which is what I suspect AI so far is, although like I said happy to change my mind on this if warranted --- from what we'd unequivocally would say is original, even if there's some trace derivative elements.

Like that story of yours? If we find that AI has cobbled together bits and pieces from like 50 different stories, or even 100 different stories, and simply pieced them together coherently enough without adding any completely novel touches: well that's original but derivative. A truly original human story on the other hand, while sure there may be some elements that are derivative, but the core of it is ...well, not derivative. Do you think AI is capable of such? Either one AI, or a bunch of as many AI bots as needed working together?

If the short answer to the paragraph immediately above is a Yes --- and if you're sure you know this, as opposed to that being simply an impression (like my own view is) --- well then, in that case, sure, AI can indeed do a Roko. That's the first part of what I'm asking.


----------

And the second part of what I was asking, that's so far not been touched on yet, at least not in the posts in these last two pages: The critical thinking bit? Not all humans do this, I know, or do it always: but definitely humans are capable of critical thinking. Do you think AI is able to, as well, AI so far? Like, if there had been nothing like Roko's basilisk, nor even Pascal's Wager: and then someone out of the blue came out with Roko's basilisk and asked AI to critique it: then can the AI of today do it on its own steam, as opposed to merely aggregating what others have had to say about it?

That's the second part of what I'd been wondering. That is, I was of the firm impression that the answer is an emphatic No, so far at least. Have you any thoughts about this?

 

[Chanakya]

LLMs certainly can do logical conclusions which are not part of their training data.

Ah. I'd thought different.

Are you sure of this? That it doesn't just mash together things? I mean, if it indeed does what you say, then that would mean that any "critical thinking" us humans can do, AI has the ability to do as well, already.

If that's true, already, then that changes how I look on AI. If that's truly the case, then we'd not have any grounds to ever reject anything on grounds that it is "merely" AI slop! If that's actually true, then we should then treat AI no different than human thoughts, and evaluate it on its actual content, always. No reason then to discount something as "just AI", as some of us tend to, as opposed to evaluating the content no less than if it came from a human.

 

[Dr.Sid]

Ah. I'd thought different.

Are you sure of this? That it doesn't just mash together things? I mean, if it indeed does what you say, then that would mean that any "critical thinking" us humans can do, AI has the ability to do as well, already.

If that's true, already, then that changes how I look on AI. If that's truly the case, then we'd not have any grounds to ever reject anything on grounds that it is "merely" AI slop! If that's actually true, then we should then treat AI no different than human thoughts, and evaluate it on its actual content, always. No reason then to discount something as "just AI", as some of us tend to, as opposed to evaluating the content no less than if it came from a human.

You just have to get some practice with it. To know what it can do, and what it can't. There practically always some value, and some nonsense. Even rephrasing your request can be useful. Overview of how your problem is typically solves might be priceless. Might also contain some complete nonsense. IMHO talking to AIs is stimulating and exciting. That must mean something. Just take every claim with grain of salt. Or rather, a shovel.

 

[Puppycow]

This is professor David Kipping from Columbia talking about a meeting he attended at the Institute for Advanced Studies at Princeton:


I'm just 10 minutes into it myself, but it seems very interesting.

 

[Dr.Sid]

Some tl;dr ?

 

[Puppycow]

It's about how top people in astrophysics are using AI to do science.
These are the topics:

CHAPTERS 0:00 Teaser0:36 My IAS Visit1:30 IAS Context2:27 Why This Setting Matters3:38 Meeting Brief4:47 Coding Supremacy6:55 Analytic Supremacy10:15 Surrendering Control12:31 Accelerating Adoption13:40 Ethic Be Damned15:05 Skill Atrophy17:01 Use It Else You're Cooked18:47 Learning to Use AI20:41 Human Oversight23:40 Cost Concerns26:47 Patents & IP29:38 Who Wins?34:04 Who Loses?40:26 Grad Admissions45:44 Collaborations48:40 Tenured Faculty51:58 Me & AI58:43 Public Backlash1:04:41 Historical Significance1:05:20 Democratization of Science1:09:00 Science as a Human Endeavour1:12:15 Final Thoughts1:14:20 Credits

I finished watching it. I'm not going to summarize further. I suggest you start watching it, and if you feel that it's a waste of your time, then stop.

 

[Chanakya]

You just have to get some practice with it. To know what it can do, and what it can't. There practically always some value, and some nonsense. Even rephrasing your request can be useful. Overview of how your problem is typically solves might be priceless. Might also contain some complete nonsense. IMHO talking to AIs is stimulating and exciting. That must mean something. Just take every claim with grain of salt. Or rather, a shovel.

Well sure, but then that applies to what humans say and write as well, right? Whole shovels of salt are definitely needed when engaging with some humans as well!

My point is, I was under the impression that AI is so far unable to ...to put it in your terms, do anything other than parody its training material. My impression was that so far, AI is incapable of either "creating" anything that is not a mashup of its training material, or of reasoning it's way to anything that is not derivative of its training material. ...If my impression was mistaken, if AI is able to do both, already, and to do it coherently not like a monkey throwing random darts: well then, if that's actually the case, then we'd have no reason to dismiss AI generated content simply because it is AI generated as opposed to human generated, would we?

Someone brought in the meme, AI;DR, earlier in the thread. If what you're saying is correct, as far as critical thinking by AI; and if what @Darat said is correct, as far as creativity from AI: well then, in that case an "AI;DR" attitude would be no more than a bias, an attitude previously valid but now outdated and fallacious, wouldn't it?

 

[Dr.Sid]

Well sure, but then that applies to what humans say and write as well, right? Whole shovels of salt are definitely needed when engaging with some humans as well!

My point is, I was under the impression that AI is so far unable to ...to put it in your terms, do anything other than parody its training material. My impression was that so far, AI is incapable of either "creating" anything that is not a mashup of its training material, or of reasoning it's way to anything that is not derivative of its training material. ...If my impression was mistaken, if AI is able to do both, already, and to do it coherently not like a monkey throwing random darts: well then, if that's actually the case, then we'd have no reason to dismiss AI generated content simply because it is AI generated as opposed to human generated, would we?

Someone brought in the meme, AI;DR, earlier in the thread. If what you're saying is correct, as far as critical thinking by AI; and if what @Darat said is correct, as far as creativity from AI: well then, in that case an "AI;DR" attitude would be no more than a bias, an attitude previously valid but now outdated and fallacious, wouldn't it?

I don't understand the duality. It's not perfect humans vs. useless AI. It's a gradual scale, AI is better at some things, people are better at other things. It depends. To dismiss anything written by AI as nonsense is shortsighted. Depends what you want, how you prompt it, and how you interpret it.
If it is about originality, LLMs can be original. They can do original code, they can do original poetry.

 

[Darat]

Well sure, but then that applies to what humans say and write as well, right? Whole shovels of salt are definitely needed when engaging with some humans as well!

My point is,
I was under the impression that AI is so far unable to ...to put it in your terms, do anything other than parody its training material. My impression was that so far, AI is incapable of either "creating" anything that is not a mashup of its training material, or of reasoning it's way to anything that is not derivative of its training material. ...If my impression was mistaken, if AI is able to do both, already, and to do it coherently not like a monkey throwing random darts: well then, if that's actually the case, then we'd have no reason to dismiss AI generated content simply because it is AI generated as opposed to human generated, would we?

...snip....

Do we have evidence that humans do anything different? We know the processes that cause the outputs of humans and AIs are different but are they both producing the same type of output? (By that I mean is it like we know the processes i.e. the types of locomotion kangaroos and snakes use to go from A to B are different but we know they both get to B.)

 

[JayUtah]

Do we have evidence that humans do anything different?

I have anecdotal observation to that effect. A lot of what we do as designers is iterative, and thus falls within the ambit of AI that can reformulate and extrapolate from training data. The result is essentially the same thing as you get from AI. We stay within a vocabulary that we know works.

However I have been privy in my career to some truly outside-the-box thinking in various fields of engineering. When these occur, they are transformative and (in our case) uniquely profitable.

Humans can think outside the box, but don't often need to. AI so far cannot seem to think outside the box. Insofar as what you need to do is inside the box, AI can be helpful.

 

[Skeptical Greg]

AI Doesn't have lightbulb moments?

 

[Chanakya]

I don't understand the duality. It's not perfect humans vs. useless AI. It's a gradual scale, AI is better at some things, people are better at other things. It depends. T
o dismiss anything written by AI as nonsense is shortsighted.Depends what you want, how you prompt it, and how you interpret it.

Well, to dismiss anything outright can, arguably, shortsighted. And I've no doubt, never had any doubt, that in theory, and probably even in practice soon one day, AI will be able to think critically, like us humans do. But so far at least, it was my understanding --- which I suppose may, just perhaps, have been flawed all along --- that all AI could do is just regurgitate different permutations and combinations of what it's already been trained on. ...So that, for instance, take Roko's basilisk that we'd been discussing: in a world where Roko's basilisk has already been analyzed (and dismissed) to death, as it has been in ours, then AI can probably present an excellent critique of it; but in a world where nothing like that thought experiment has ever been produced by anyone, then AI will draw a complete blank, and fall back on talking nonsense. (Which is what many humans do as well, but at least us humans do have the capacity for critical thinking, sometimes, and to use our critical thinking as an engine for arriving at truth, at as close an understanding of reality as is possible. It was my understanding that that is something AI could not do.)

It seems to be to be the case that (you're saying that) AI has now already crossed over onto the stage where it can actually do critical thinking on its own steam. Where it might be able to critique Roko's baslisk from different angles, and do it as well as a human might, even in a world where nothing like that construct has ever been presented and where it encounters this idea for the very first time.

If that is correct ---- and I'll take it for now that it's correct. unless someone comes across and persuasively contests this view ---- then ...well, I guess it's time for me to update how I look on AI "slop". There's no more reason, then, to think of AI as a producer of slop than are humans (since us humans do that too, often enough).

If it is about originality, LLMs can be original. They can do original code, they can do original poetry.

Yep, apparently. That is, I was under the impression that is not the case, so far. But it seems I was mistaken in thinking that we're not there yet. Apparently we are. ...Well then, it looks like I need to update how look at AI as far as that too.

So that, in terms of both original creative ideation, as well as critical thinking, AI is already on par with us. (Of course, as you say, in some specific areas we might be better still, but even that is surely just a matter of time, and not very long at that.)

 

[JayUtah]

AI Doesn't have lightbulb moments?

Not in my experience. The transformational moments I’ve seen stem largely from questioning core assumptions. Mathematically this is not something an LLM can do. An LLM’s notion of applicability is mathematically (specifically a cosine) welded to the notion of conformity. If the training data assumes the core assumption too, it will simply be there underneath the data, but not accessible as a separate computable object.

 

[Chanakya]

Do we have evidence that humans do anything different? We know the processes that cause the outputs of humans and AIs are different but are they both producing the same type of output? (By that I mean is it like we know the processes i.e. the types of locomotion kangaroos and snakes use to go from A to B are different but we know they both get to B.)

Well, we've so far acted as if that were so, isn't it. ...Not to make this about FMF, not talking about what we should do or not do here, but addressing the principle of what we did (a principle I was in agreement with): we'd discourage people, actually bar people, from posting large portions of AI material, right? ...Now why would we do that? Say we're discussing some subject substantially. Then, if it is the case that AI is indeed able to think critically (as opposed to merely mash together what it's seen), then we'd be very ill advised to not allow into the discussion perhaps some unique meaningful POV.

Like I said, I was in agreement with that policy. And that is because, in general and without necessarily bringing our forum into it, that's how I myself conducted my own research as well, and what I viewed as the right approach generally: that for anything that is at all important, one might start out with AI if one wishes, but then one goes on to check up its links, and to engage directly with the source material it's used.

But now I see that AI is already able to produce critical thinking beyond the sources it tracks. So that we'd be potentially missing out on some fresh and meaningful perspective unless we started engaging with AI ---- over and above looking up sources, over and above discussion with humans.

So that's something that has changed, it would seem. Changed already. That future's here, already. Or so I take you to be saying here, as far as AI and critical thinking and original ideation.

 

[Chanakya]

Not in my experience. The transformational moments I’ve seen stem largely from questioning core assumptions. Mathematically this is not something an LLM can do. An LLM’s notion of applicability is mathematically (specifically a cosine) welded to the notion of conformity. If the training data assumes the core assumption too, it will simply be there underneath the data, but not accessible as a separate computable object.

Can you explain that in simpler terms?

(Didn't quite follow your explanation, how you've worded it. But if AI is already capable of original ideation and critical thinking, which is what is suggested here in these recent posts, then why might that not extend to examining those assumptions as well, and/or looking at scenarios where those assumptions do not apply?)

I mean, if AI today is capable of original ideation of say Roko's basilisk, and original critical thinking/evaluation of that construct, then why, in principle, might it not also be capable of original ideation of a heliocentic system, or relativity, as well as critical thinking/evaluation of such?

 

[Beelzebuddy]

Creativity and critical thinking are very different things. LLMs fail utterly at the latter.

Seriously just try some out for yourself. AI's getting crammed into everydamnthing. Every browser and search engine has its own and is desperate for someone to use it.

 

[JayUtah]

Can you explain that in simpler terms?

No, only in more complicated terms.

(Didn't quite follow your explanation, how you've worded it. But if AI is already capable of original ideation and critical thinking, which is what is suggested here in these recent posts, then why might that not extend to examining those assumptions as well, and/or looking at scenarios where those assumptions do not apply?)

LLMs work by encoding semantic content as tensors. Not all semantic content is explicit enough to be separately representable along its own tensor dimension, or some recognizable combination of them. This results in embeddings that embody the underlying assumptions without actually expressing them in the embedding by something that is exposed to computation. You can certainly generate contrary content, but the transformer layers of the LLM will reject it according to probabilistic rules.

Imagine you have a training data for a baking AI that's composed of millions of baked-goods recipes. You have a variety of ingredients, methods, and baking regimes. An AI can synthesize a number of variation on the cookie from this, all of which have various probabilistic landscapes in the overall problem space. They're all credible (although actual cooks note that AI-generated recipes tend to do stupid things).

But if some human baker devises an unprecedented method whereby the dough is actually frozen instead of baked, and the desired chemical reactions occur, then such a thing would simply not be visible in any way in the training data. The training data embeds such notions as chemical composition of the dough, baking temperature, baking time, oven humidity, and so forth. And this can be further enriched with metadata on which combinations have produced good outcomes. But what is not embedded in the data is the decision to use an endothermic process at all. That's simply assumed in all the training data and therefore not represented as a differentiable concept. Hence it will be assumed in all the solutions.

None of that really has anything to do with the higher-level philosophical discussion of such things as Roku's Basilisk.

 

[arthwollipot]

I finished watching it. I'm not going to summarize further. I suggest you start watching it, and if you feel that it's a waste of your time, then stop.

I'm afraid I'm going to need a whole lot more than that before deciding to commit to watching a video that's over an hour long.

 

[Puppycow]

I'm afraid I'm going to need a whole lot more than that before deciding to commit to watching a video that's over an hour long.

Who says you need to "commit" to watching it? Like I said, as soon as you start to feel that it's a waste of your time, you can stop at any time.

ETA: I watched the whole thing and thought it was well worth it.

 

[arthwollipot]

Who says you need to "commit" to watching it? Like I said, as soon as you start to feel that it's a waste of your time, you can stop at any time.

I feel that it's a waste of my time now, even before starting it.

ETA: I watched the whole thing and thought it was well worth it.

Okay, good for you. That's not a reason for me to watch it.