AI and science and art

[discovolante]

Sorry in advance, I had a massive cup of coffee this morning and forgot that I'm pretty sensitive to it now, so I've got the jitters.

Anyway I'm sorry for this thread too because even though I started it, I'm getting slightly tired of the obsessing over Chat GPT and DALL-E 2 and all that. But I've listened to a few discussions about it anyway, from tech and arty and philosophical perspectives, with people who specialise in each of those areas. And I feel like each time, they find themselves bumping up against their ignorance of the other disciplines. Artists don't fully understand what the AI does or what it might or might not be capable of, scientists/tech people don't really understand what art is, philosophers are good at asking questions of both but can't really answer any of them. One thing all of them seem to grapple with unsuccessfully is the question of what consciousness/sentience actually is. I may be a bit behind here but I feel like until now there's been a kind of underlying assumption that it's ultimately an ability to be aware of your own existence, or think about yourself. But listening to these discussions it does seem to be that creativity is at the root of it, because we seem no more capable of being able to properly identify what creativity is than what consciousness is. I think the artists probably have a better grasp of it, as they understand the process more, but they also have more of a vested interest in identifying it as something uniquely erm, not human but biological I suppose.

Possibly I seem more stoned than high on caffeine but there you are.

[Brightonian]

Trying out ChatGPT etc. is on my to-do list, especially after this effusive video from Tom Scott.

Or maybe it's all a con, and these things are really a massive number of real people in an office block on the outskirts of Manila.

[nekomatic]

Artists don't fully understand what the AI does or what it might or might not be capable of, scientists/tech people don't really understand what art is, philosophers are good at asking questions of both but can't really answer any of them.

A dilettante in all three fields writes…

The obvious comment on these AI models is that they aren’t intelligent, conscious or creative because all they do is produce mashups of pre-existing human-created content in response to their inputs, thus they are simply tools or techniques akin to sampling or collaging, just in a more superficially sophisticated way.

The obvious riposte to that is to ask whether human artists actually do anything fundamentally different, give or take the occasional step-change disruption of a Picasso or a Miles Davis.

I expect that philosophers can give us multiple possible definitions of consciousness, and for each one we could answer fairly well whether the AI meets it or not, but not a verdict on which definition is actually correct.

I guess my take on it - which I’d welcome philosophical comment on - is that if we accept that consciousness is ultimately a property of our brains which are physical systems, and not some spiritual or metaphysical woo thing, then there’s no fundamental reason that a sufficiently complex and powerful computer model that’s capable of modifying itself over time can’t achieve what we would define as consciousness. That doesn’t mean ChatGPT is anything near there though.

[dyqik]

Trying out ChatGPT etc. is on my to-do list, especially after this effusive video from Tom Scott.

Or maybe it's all a con, and these things are really a massive number of real people in an office block on the outskirts of Manila.

If those people have limited English, then you're in the Searle's SE Asia Chinese room scenario.

[discovolante]

Trying out ChatGPT etc. is on my to-do list, especially after this effusive video from Tom Scott.

Or maybe it's all a con, and these things are really a massive number of real people in an office block on the outskirts of Manila.

I've had a go at it and asked its views on stuff and its responses could probably be best described as 'undergraduate level' - reasonably considered and seeming to 'understand' the subject matter but no original thought or ability to form a conclusion of its own.

[discovolante]

Artists don't fully understand what the AI does or what it might or might not be capable of, scientists/tech people don't really understand what art is, philosophers are good at asking questions of both but can't really answer any of them.

A dilettante in all three fields writes…

The obvious comment on these AI models is that they aren’t intelligent, conscious or creative because all they do is produce mashups of pre-existing human-created content in response to their inputs, thus they are simply tools or techniques akin to sampling or collaging, just in a more superficially sophisticated way.

The obvious riposte to that is to ask whether human artists actually do anything fundamentally different, give or take the occasional step-change disruption of a Picasso or a Miles Davis.

I think they/we do. In fact this came up in the interview I was listening to that prompted this thread (an interview with Adrian Tchaikovsky). They discussed it generally but he also gave an example of the Transformers TV series and how repetitive it is, with basically the same plot each time but with a different villain. I thought that was a pretty poor example because 1) it's for children who tend to like repetition and 2) it's an almost entirely commercial venture with no real interest in actual creativity.

I think one person (on the arty side of things, very much not an AI expert) described the difference quite well. They said that you could ask a person to draw a picture of something but what they end up drawing could be something only loosely connected with what you've asked them to do because they have the ability to let their imagination take them off in different directions.

Of course lots of stuff can be derivative of or influenced by stuff that came before, but that doesn't mean it's nothing more than a kind of collage.The same facial features, or even groups of facial features, can appear on different people, but that doesn't mean that any two faces will be the same.

This sort of reminds me of the Borges short story where someone tries to write Don Quixote word for word, but not actually copy it, heh.

I expect that philosophers can give us multiple possible definitions of consciousness, and for each one we could answer fairly well whether the AI meets it or not, but not a verdict on which definition is actually correct.

I guess my take on it - which I’d welcome philosophical comment on - is that if we accept that consciousness is ultimately a property of our brains which are physical systems, and not some spiritual or metaphysical woo thing, then there’s no fundamental reason that a sufficiently complex and powerful computer model that’s capable of modifying itself over time can’t achieve what we would define as consciousness. That doesn’t mean ChatGPT is anything near there though.

Yeah it's a difficult one to escape isn't it, although I suppose one fairly happy conclusion from it is that we don't necessarily need to see other animals as fundamentally different from us - obviously not something to get carried away by but I think it makes life a bit more interesting.

[nekomatic]

you could ask a person to draw a picture of something but what they end up drawing could be something only loosely connected with what you've asked them to do because they have the ability to let their imagination take them off in different directions.

Well, I don’t think we can exclude that a big computer model could do that, if its parameters are adjusted to let that happen. Is imagination anything more than the ability to draw conclusions that are further than usual away from our inputs?

[Woodchopper]

you could ask a person to draw a picture of something but what they end up drawing could be something only loosely connected with what you've asked them to do because they have the ability to let their imagination take them off in different directions.

Well, I don’t think we can exclude that a big computer model could do that, if its parameters are adjusted to let that happen. Is imagination anything more than the ability to draw conclusions that are further than usual away from our inputs?

I spent a bit of time in December playing around with one of the AI image generators. Just enter some keywords and it produces an image. It was doing that, coming up with takes on the keywords that I hadn't thought of.

I'm not suggesting that this is the same as an artist being creative. I think that Nick Cave has the last word there.

But what is happening is that the realm of things that humans can do that computers can't keeps getting smaller. Its incremental steps and with each the humans have to redefine ways that they are unique. I have no idea where the process will end. Despite its effective mimicry Chat GPT is a very long way from replicating human intelligence.

[Woodchopper]

I guess my take on it - which I’d welcome philosophical comment on - is that if we accept that consciousness is ultimately a property of our brains which are physical systems, and not some spiritual or metaphysical woo thing, then there’s no fundamental reason that a sufficiently complex and powerful computer model that’s capable of modifying itself over time can’t achieve what we would define as consciousness. That doesn’t mean ChatGPT is anything near there though.

Maybe, or at least it'll become harder and harder to tell the difference.

But our brains are built very differently from any machine yet devised. So what we think of as consciousness could be a property of the specific way that our meat computers are made.

[Brightonian]

Trying out ChatGPT etc. is on my to-do list, especially after this effusive video from Tom Scott.

Or maybe it's all a con, and these things are really a massive number of real people in an office block on the outskirts of Manila.

This AI startup that "generated" 3D models from 2D art was actually using human workers

[Brightonian]

Has anyone used any AI tools to produce programming code? I'm wondering whether I should even bother. GPT-4 is starting to be downright annoying.

I've just spent about three hours trying to get it to generate code that won't fall over. I wrote the requirement in just 15-20 minutes. Without even having to clarify my requirement, I went through 12 rounds of:

Here's your code <etc.>

No, got error message <etc.>

On the 13th round it finally worked.

Sometimes it's just me, but usually I can express my requirements clearly, and it's GPT-4 that's away with the fairies. This time it really took the biscuit. But that's fine - I really didn't want to go sitting outside the pub with a beer.

[bjn]

Not going there. Partly on principle, because if they want to use my data to train their LLM, whether my requirements or my code, they can pay me. Partly because what I tend to code isn’t sufficiently with in their training set to generate anything useful (I am special me). Partly because the code generated isn’t particularly good anyway.

[dyqik]

Has anyone used any AI tools to produce programming code? I'm wondering whether I should even bother. GPT-4 is starting to be downright annoying.

It's a known route for introducing security exploits into code, particularly via libraries that GPT makes up, and which hackers then create to bundle exploits onto the host machine.

[Imrael]

Has anyone used any AI tools to produce programming code? I'm wondering whether I should even bother. GPT-4 is starting to be downright annoying.

I messed around with getting solutions to a database admin technical issue. The result was a process of legal correctly described steps that might have solved the problem. But failed to take account of some normal factors in a production environment and could have been actively destructive.

The issue here I think is that an inexperienced person could try to follow advice like this and cause havoc.

[snoozeofreason]

I've tried getting ChatGPT to solve cryptic crossword clues, and found the results quite revealing. For example given the clue

Police document conflict in NHS, rat race ghastly (6,7)

The answer to this cryptic crossword clue is "MINUTE BOOK"

Explanation:

"Police document" refers to a record kept by the police,
"Conflict" could be a homophone for "clash"
"NHS" is National Health Service, a government-funded healthcare system in the UK
"Rat race" could be a metaphor for a busy and hectic lifestyle
"Ghastly" could mean something is unpleasant, or could also be a homophone for "lastly"

So "Record kept by the police of a clash in the National Health Service, Busy and hectic lifestyle lastly" is "MINUTE BOOK" (6,7) means the answer is two words and the length of the first word is 6 letters and the second word is 7 letters.

The interesting thing about it is that the explanation, on a superficial level, seems to display an understanding of how cryptic crosswords work. It makes it look as if ChatGPT knows that homophones and metaphors are relevant and that the numbers in brackets refer to the lengths of the words in the clues. But then it provides an answer that has no relation to the explanation it has just given (not even getting the lengths of words correct).

(For non crosswordy types, the real answer is "Search Warrant" - anagram of "NHS, rat race" placed around the word "war").

[Rich Scopie]

I think one person (on the arty side of things, very much not an AI expert) described the difference quite well. They said that you could ask a person to draw a picture of something but what they end up drawing could be something only loosely connected with what you've asked them to do because they have the ability to let their imagination take them off in different directions.

We did this in one if those pointless meetings that HR likes to schedule you in for. We were all asked to draw a flying horse.

Lots of crap depictions of Pegasus, a few horses held up by balloons etc, but I was the only one who drew a pub.

I suspect that says more about me than HR really needed to know.

[shpalman]

I think one person (on the arty side of things, very much not an AI expert) described the difference quite well. They said that you could ask a person to draw a picture of something but what they end up drawing could be something only loosely connected with what you've asked them to do because they have the ability to let their imagination take them off in different directions.

We did this in one if those pointless meetings that HR likes to schedule you in for. We were all asked to draw a flying horse.

Lots of crap depictions of Pegasus, a few horses held up by balloons etc, but I was the only one who drew a pub.

I suspect that says more about me than HR really needed to know.

flying horses

[Stranger Mouse]

Nice Daily Skit here. Worth watching all the way through Big Boy

https://x.com/thedailyshow/status/17907 ... 28083?s=61

[hakwright]

Has anyone used any AI tools to produce programming code? I'm wondering whether I should even bother.

You shouldn't

Despite the huge advances AI has made in generating text, images, music etc - it is still (as described earlier) nothing more than a very effective mimic. Any task that requires a level of understanding to be successful is likely to fail pretty badly (although it may look/sound impressive to the casual observer).

I've just spent about three hours trying to get it to generate code that won't fall over. I wrote the requirement in just 15-20 minutes. Without even having to clarify my requirement, I went through 12 rounds of:

Here's your code <etc.>

No, got error message <etc.>

On the 13th round it finally worked.

There's an interesting AI Weirdness post on using AI to generate images, then trying to make small alterations to the output with multiple feedback iterations. The blog post title is "An exercise in frustration".

https://www.aiweirdness.com/an-exercise-in-frustration/

I imagine trying to get functioning code this way is likely to be an order of magnitude more frustrating than just trying to generate a single image.

I don't believe the work done by software engineers, journalists, artists and musicians will be seriously impacted by novices using AI. But there will almost certainly be a trend for people already skilled in those areas using AI in novel ways to enhance their capabilities.

[Woodchopper]

Has anyone used any AI tools to produce programming code? I'm wondering whether I should even bother.

You shouldn't

Despite the huge advances AI has made in generating text, images, music etc - it is still (as described earlier) nothing more than a very effective mimic. Any task that requires a level of understanding to be successful is likely to fail pretty badly (although it may look/sound impressive to the casual observer).

Well yes, and no. I agree that it doesn't understand anything in the sense that you or I do.

But I have been impressed and a bit disturbed a few times. I've found it useful to upload circa 5000 words of text and ask ChatGPT to summarize the key points in the text I've given it. It often highlights relevant things that I hadn't thought of (and I've got decades of experience and a PhD etc). It seems to work quite well doing a task that previously I'd think of a being something that only an experienced human editor could do. Of course GPT is just identifying patterns, but ultimately I don't care.

[IvanV]

...Of course GPT is just identifying patterns...

Codifying the laws of nature is really no more than identifying patterns. Our present established laws of nature are exceedingly accurate in many regards, but fail to account for a number of inconvenient facts about the universe, and infamously make some predictions that are wrong by very many orders of magnitude. If AI is so good at identifying patterns, perhaps it could propose some better codifications of the laws of nature that are more consistent with the known facts.

[dyqik]

...Of course GPT is just identifying patterns...

Codifying the laws of nature is really no more than identifying patterns.

It's a lot more than identifying patterns. If it was just identifying patterns, then epicycles would still be dominant version of astronomy. There'd be no Newton. Codifying the laws of nature requires physical insight and intelligence beyond just identifying patterns. Science explains the patterns. Just recording them is stamp-collecting.

If AI is so good at identifying patterns, perhaps it could propose some better codifications of the laws of nature that are more consistent with the known facts.

No, current "AI" can only identify and repeat patterns. It cannot form intelligent thoughts about the patterns and why the patterns exist.

ETA: the AI model of pattern identification really is just adding more and more epicycles to a bad model. Which is why it's so hideously inefficient and a significant contributor to emissions.

[IvanV]

...Of course GPT is just identifying patterns...

Codifying the laws of nature is really no more than identifying patterns.

It's a lot more than identifying patterns. If it was just identifying patterns, then epicycles would still be dominant version of astronomy. There'd be no Newton. Codifying the laws of nature requires physical insight and intelligence beyond just identifying patterns. Science explains the patterns. Just recording them is stamp-collecting.

If AI is so good at identifying patterns, perhaps it could propose some better codifications of the laws of nature that are more consistent with the known facts.

No, current "AI" can only identify and repeat patterns. It cannot form intelligent thoughts about the patterns and why the patterns exist.

ETA: the AI model of pattern identification really is just adding more and more epicycles to a bad model. Which is why it's so hideously inefficient and a significant contributor to emissions.

I was perhaps being provocative. There is a sense in which Newton's law is a pattern generator. But you are saying that AI, as it currently exists, only recognises surface patterns - the epicycles - it can't look for the deep pattern generator, that might explain multiple phenomena, such as apples falling off trees as well as astronomical observations.

But is this because we are thinking of AI as LLMs? There are other kinds of AI which are trained to do other, more specific, things, like analyse data, play games, etc. AI that does things like play chess, etc, tests large numbers of options, complying with the laws of the game, and finds the one that best meets an objective. Originally computer chess programmes did this by brute force, but AI is more effective, because the branching of options is so large and AI cuts down the options you test, so enabling deeper useful analysis with less calculation. I was wondering if, for example, this might be a method of trying to prove mathematical hypotheses, if it could be trained to recognise valid proofs, and generate valid proof steps, and so search for a route to a complete valid proof of the hypothesis. That would appear to be an easier problem than looking for physical laws, but it is a step on the way.

[dyqik]

AI, of any variety, doesn't postulate new models based on conceptual understanding. It can only form patterns from it's existing concepts.

This makes it useless for science.

[monkey]

They've tried getting of AI to do physics at least once. I remember this story from a couple of years back - clicky.

They got a neural net to model dynamic systems like a double pendulums of various types. The neural net came up with models that worked - it identified patterns, worked out how to describe them in maths, and made good predictions.

The trouble was, there was no explanation of what the variables mean, and in meat physics variables have meaning so's you know what's going on. So the researchers had no idea if what was is doing is useful or not. All they knew is that it was doing things differently to how a meat physicist would do it.

I think this was one of Dyqik's points.

(link to actual paper - clicky)