Merged Artificial Intelligence

[Chanakya]

An LLM-based generative AI utterly fails at critical thinking because evaluating concepts according to some criterion extrinsic to its training data is just not something the algorithm is even aimed at doing. However, if critical analysis of a concept appears as part of the training data, then a generative AI can synthesize what looks like a critical analysis if prompted to do so. But it won't be an independent analysis of the data. It will just be the critical-analysis part of whatever data it was fed, transformed as appropriate to the prompt.

Yep. Like I said, this aligns completely with where I was at myself.

And yet it's been said here, in recent posts, that these AI thingies are actually able to critical-think their way to conclusions that are not part of their training material. That's either true, in which case I need to update my views, already; else it is not true, in which case, well, we remain where I'd thought we're at.

(Heh, in this respect I'm afraid my appreciation of this subject is a bit like AI's! In the sense that, lacking the technical grounding of how the innards of AI actually works, and of where AI models have actually reached, in terms of what happens in the inside of them: then I'm not myself capable of reliably critical-thinking my way to the conclusion of the question I'm asking. So that, much like AI, on this subject I find myself in the position of AI models, having to base my own conclusion on others' critical appreciation of this, others who are better versed technically on this than I am.)

As to whether a generative AI can be creative, I think that's still mired in the details of what actually constitutes creativity.

Can AI come up with Roko's Basilisk in a world where nothing like it has ever been produced? (Well Pascal's Wager maybe, but not more directly similar themes, like for instance the SF that @The Great Zaganza referenced.) ([eta]Or, hell: Can AI come up with Pascal's Wager, in a world where Pascal never thought it up?[/eta]) ...Can AI come up with the idea of a heliocentric system in a world where such has never ever been proposed? ...Can AI come up with Einstein's ideas about gravity in a world where Newton is all there is.

Like that. At least that's what I was asking about.

 

[The Great Zaganza]

LLM constantly come up with new Chess Moves that the Rules never envisioned ....

 

[Wudang]

One problem is that "critical thinking" is perhaps too fuzzy a concept to be useful here beyond a certain point.
Yes I've had a few goes at refining this point. No I'm not happy with any of my attempts thus far.

 

[Dr.Sid]

We don't really need to define "critical thinking" .. we need to devise a test.

 

[Wudang]

We need at least some form of operational definition to devise a test.
And I still somewhat agree with whoever said that the Turning test is more a test of our ability to anthropomorphise things.

 

[Skeptical Greg]

LLM constantly come up with new Chess Moves that the Rules never envisioned ....

Do you have any further context for this?

The rules are well defined and pretty simple overall.

 

[W.D.Clinger]

If someone actually knows, not just vaguely but with certainty, basis their familiarity with AI and how it actually works, then I wish they'd just say it. I really don't see how I can ask any more clearly than this. Answer not with would-be wise hints, but with a clear answer to a clear question.

I think part of your problem here is that none of us are software engineers with a specialisation in AI programming.

Part of my problem is that, long ago, I took quite a few graduate-level courses in artificial intelligence and became moderately familiar with the AI of that era. Afterwards, I taught a course or two myself on AI or AI programming. Then, for several decades, I attended several presentations a year covering recent research on AI, machine learning, and the like.

Part of the problem with @Chanakya's question is that the very concept of artificial intelligence is squishy and has been so from its origins.

One of the best definitions I've heard is that AI is something computers do that, if done by a human, would be regarded as evidence of intelligence but, if done by computers, will not be regarded as evidence of intelligence once you understand how the computers are doing it.

There was a time when doing arithmetic was regarded as evidence of intelligence. When computers began to do arithmetic better and faster than any human, doing arithmetic was no longer regarded as evidence of intelligence.

There was a time when doing calculus integrals was regarded as evidence of intelligence. Then James Slagle came along. If I recall correctly, there is an algorithm that solves the problem as well as it can be solved by any human or computer, and that algorithm has been implemented. So doing integrals now has the status of doing arithmetic.

Once upon a time, Alan Turing proposed that when a computer, in written conversation, could fool you into thinking it was human, it should be regarded as intelligent. Then Joseph Weizenbaum came along, and showed how easy it was to fool a lot of people. Much of this thread has been devoted to the progress that has been made toward fooling even more people more reliably, so I needn't say any more about that.

There was a time when playing chess was regarded as evidence of intelligence. Computers began to play serious chess in the 1960s. By 2000, the best computer chess programs were competitive with grandmasters. The best programs now routinely defeat the world's best humans.

Language translation, facial recognition, image matching, melody matching, voice recognition, speech-to-text, and autonomous vehicles have all been goals of AI research, and all of those goals have been achieved to some degree. Progress is still being made, but we understand how those things work, so they are no longer regarded as evidence of intelligence.

I could go on.

My point here is that (1) artificial intelligence, when it succeeds, comes to be taken for granted, regarded as ordinary technology instead of machine intelligence, and (2) an awful lot of successful AI has been developed.

One of the more important lessons we've learned from the history of artificial intelligence is that it's hard to beat the combination of brute force and large data. That's why AI data centers consume so much electricity, and it's why there is such a rush to build even larger and more resource-intensive centers.

But as for how AI actually works, there really isn't a whole lot of mystery about that. I am no expert, but give me a specific example of something you think of as AI, and I can probably give you some idea of how it works.

 

[The Great Zaganza]

Do you have any further context for this?

The rules are well defined and pretty simple overall.

LLMs like GPT routinely make illegal moves in Chess - because they don't understand the rules and will just extrapolate from past matches.

 

[Darat]

LLM constantly come up with new Chess Moves that the Rules never envisioned ....

And the same with Go.

Plus as ever we need to remember now that there is more to the likes of ChatGPT than just the LLM backbone.

 

[Darat]

Yep. Like I said, this aligns completely with where I was at myself.

And yet it's been said here, in recent posts, that these AI thingies are actually able to critical-think their way to conclusions that are not part of their training material. That's either true, in which case I need to update my views, already; else it is not true, in which case, well, we remain where I'd thought we're at.

(Heh, in this respect I'm afraid my appreciation of this subject is a bit like AI's! In the sense that, lacking the technical grounding of how the innards of AI actually works, and of where AI models have actually reached, in terms of what happens in the inside of them: then I'm not myself capable of reliably critical-thinking my way to the conclusion of the question I'm asking. So that, much like AI, on this subject I find myself in the position of AI models, having to base my own conclusion on others' critical appreciation of this, others who are better versed technically on this than I am.)





Can AI come up with Roko's Basilisk in a world where nothing like it has ever been produced? (Well Pascal's Wager maybe, but not more directly similar themes, like for instance the SF that @The Great Zaganza referenced.) ([eta]Or, hell: Can AI come up with Pascal's Wager, in a world where Pascal never thought it up?[/eta]) ...Can AI come up with the idea of a heliocentric system in a world where such has never ever been proposed? ...Can AI come up with Einstein's ideas about gravity in a world where Newton is all there is.

Like that. At least that's what I was asking about.

Is the critique of the video not an original critique? We know that video is not in its training data, nor was there any critique of that video in its training data.

 

[JayUtah]

LLM constantly come up with new Chess Moves that the Rules never envisioned ....

I don't think you meant to reference the rules. Moves outside the rules are illegal. I suspect you meant to say they were moves outside the conventional gameplay. To some extent, chess is a formulaic game.

Are they winning moves? Or is the phenomenon simply that they don't follow the traditional openings, gambits, etc.? We've been teaching computers specifically to play chess for decades.

The domain of chess moves is both unambiguously defined and strictly bounded. Show me any chess board with any allowable combination of pieces and I can completely enumerate literally all the possible next moves. And from each of those I can enumerate all the other side's possible following moves. Because the solution set is closed and bounded, it's not the same as an open-ended prompt space of training space. So it's comparatively simple to generate chess moves that fall outside convention, but it's not at all like navigating an open-ended tensor space where convention is baked in.

 

[Skeptical Greg]

A game of chess should not proceed beyond an illegal move.

 

[The Great Zaganza]

I was just riffing on Chanakya's doubts about LLMs being able to come up with something new - when fed with the correct and only rules of Chess, a LLM will nevertheless come up with new, illegal moves that it thinks will be in line with the game rules - because it doesn't understand what a rule is.

 

[arthwollipot]

Can AI come up with Roko's Basilisk in a world where nothing like it has ever been produced?

I think you're too hung up on the idea that Roko's Basilisk was a wild and crazy new thing that required a quantum leap of creativity to come up with, rather than a straightforward thought experiment that someone once posted on social media on a lazy afternoon.

 

[JayUtah]

I was just riffing on Chanakya's doubts about LLMs being able to come up with something new - when fed with the correct and only rules of Chess, a LLM will nevertheless come up with new, illegal moves that it thinks will be in line with the game rules - because it doesn't understand what a rule is.

I see, so you really were talking about rules and I didn't take your meaning.

This is where you have to decide how the information will be represented. If you have a completely generalized information model, then training it on actual games will result in high probabilities for winning moves and low probabilities (if not outright zero) for illegal moves.

On the other hand, if you have a specialized representation, you can have, say, a 4096-size tensor that encodes 64 × 64 combinations of beginning and ending moves. Each move at each step then has an index, and at that index is the probability of that move being a winning move based on how you examine games. The advantage here is that you can encode the rules of chess into the trainer.

I think I see the higher genius in your proposal. If you take the initial approach with a generalized embedding, the rules won't be expressly represented in the embedding. They'll naturally fall into the space of legal moves, but only on a probabilistic basis. The rules aren't expressly represented as a separate computable entity, such that no part of the evaluation engine will say at any time, "No, that's an illegal move." It will just relegate illegal moves to such a low order of probability that they will never arise during policy synthesis—until they do.

Funny thing is that Monte Carlo methods succeed in gameplay AIs far more often than you expect. When you have policies (multiple-move strategies) that depend on the other player making predictably smart moves, a player that makes a dumb move (i.e., a randomly generated but legal one) can throw off the policy. If your policy gambles on an initially weak move that will be compensated for in a later move (a tactical weakness in favor of a strategic advantage), a random move may defeat it. No, random moves don't often win. But they win more often than you think.

 

[steenkh]

I was just riffing on Chanakya's doubts about LLMs being able to come up with something new - when fed with the correct and only rules of Chess, a LLM will nevertheless come up with new, illegal moves that it thinks will be in line with the game rules - because it doesn't understand what a rule is.

I have played with children who sometimes made illegal moves because they didn't understand the rules properly. I nevertheless regard them as intelligent beings,

 

[Darat]

I was just riffing on Chanakya's doubts about LLMs being able to come up with something new - when fed with the correct and only rules of Chess, a LLM will nevertheless come up with new, illegal moves that it thinks will be in line with the game rules - because it doesn't understand what a rule is.

Yes and no, a vanilla LLM AI could but not the current crop.

 

[The Great Zaganza]

I see, so you really were talking about rules and I didn't take your meaning.

This is where you have to decide how the information will be represented. If you have a completely generalized information model, then training it on actual games will result in high probabilities for winning moves and low probabilities (if not outright zero) for illegal moves.

On the other hand, if you have a specialized representation, you can have, say, a 4096-size tensor that encodes 64 × 64 combinations of beginning and ending moves. Each move at each step then has an index, and at that index is the probability of that move being a winning move based on how you examine games. The advantage here is that you can encode the rules of chess into the trainer.

I think I see the higher genius in your proposal. If you take the initial approach with a generalized embedding, the rules won't be expressly represented in the embedding. They'll naturally fall into the space of legal moves, but only on a probabilistic basis. The rules aren't expressly represented as a separate computable entity, such that no part of the evaluation engine will say at any time, "No, that's an illegal move." It will just relegate illegal moves to such a low order of probability that they will never arise during policy synthesis—until they do.

Funny thing is that Monte Carlo methods succeed in gameplay AIs far more often than you expect. When you have policies (multiple-move strategies) that depend on the other player making predictably smart moves, a player that makes a dumb move (i.e., a randomly generated but legal one) can throw off the policy. If your policy gambles on an initially weak move that will be compensated for in a later move (a tactical weakness in favor of a strategic advantage), a random move may defeat it. No, random moves don't often win. But they win more often than you think.

I am quite familiar with Monte Carlo and many other multi-dimensional space exploration functions - they are sampling all possible configurations a system can be in.
But there are configurations in Chess that are explicitly forbidden (like a Pawn being promoted into a Pawn or moves when in check that do not get you out of Check). Random sampling would explore such configurations and would have no way of knowing that it is in fact not possible withing the rules of the game.
And why would you need random sampling when you have a library of millions of games, and thus could put probabilities on every configuration based on their frequency of occurrence?

There are few data sets as targeted, comprehensive and accurate as full chess games, making them, supposedly, the perfect training material for a LLM - the fact that they still fail points to a fundamental flaw with the concept.

 

[The Great Zaganza]

I have played with children who sometimes made illegal moves because they didn't understand the rules properly. I nevertheless regard them as intelligent beings,

would you still think they are intelligent when they still make an illegal move after studying millions of master chess games for centuries in human years? Or would you surmise that they just can't understand the rules?

 

[Darat]

would you still think they are intelligent when they still make an illegal move after studying millions of master chess games for centuries in human years? Or would you surmise that they just can't understand the rules?

Have you an example of a LLM trained on the datasets you mentioned?