Hacker News Reader

Comment by TheOtherHobbes | Hacker News Reader

All of these examples fail on structure. The "novels" don't have a plot, the "music" doesn't have a direction, the "poem" mimics the tropes but not the rhetorical intent.

Really, NN-based art is just a more abstract kind of sampling - blundering around inside a human-originated space, bumping off the walls, looking for the occasional example that a human will find interesting.

It's still mechanical, it's still mimicry, and it's still repetitive. Even if nice things fall out - and sometimes they do - they're really just distortions of the originals. (That's particularly obvious in the library "paintings.")

They're still waiting for a human to say "I like that", but with no sense of what makes a human say that.

It reminds me of fractal art - pretty, but monotonous and disposable.

How many people can remember their favourite example of fractal art?

View on HN

The majority of human-produced art is also mechanical, mimicry and repetitive. What is the "magic" of humans that AI lacks?

This is exactly what naive AI art will reveal: that most human art is cliche ridden and derivative.

By naive AI art I mean just using it blindly to generate something that imitates X, or generating random art-like output.

There’s another potential angle to AI and art though. Humans could use AI as a tool to make art just as they have with electronic oscillators, photographic film, 3D renderers, etc. People once decried electronic music and photography as “not real art” too, so I assume some will say the same of anything using AI.

Personally I think AI as a rendering tool could be fascinating. The artist would be a little like a DJ cueing up patterns and themes, but you could get way more sophisticated than you can with mixing tracks and could do things other than music.

> What is the “magic” of humans that AI lacks?

Based on his examples,

> The "novels" don't have a plot, the "music" doesn't have a direction, the "poem" mimics the tropes but not the rhetorical intent.

AI lacks an actually understanding about what it is talking about.

Current AI is like a cargo cult I guess, they see the surface patterns but do not have the underlying understanding as to why the patterns are the way they are.

> AI lacks an actually understanding about what it is talking about.

you could claim that a chess AI lacks understanding of chess. And yet, objectively, it performs better than a human.

The current state of the art AI generated works are probably low quality, but one cannot claim that this won't improve. At some point, i would imagine that the AI generated works would be indistinguishable from human created works, even tho the AI "doesn't actually understand" it.

> you could claim that a chess AI lacks understanding of chess.

No you can't, the AI has 100% of the rules of chess, there are no rules the human understands that the AI doesn't understand. Compare that to human written texts, we haven't been able to code all the rules for human written text so the AI doesn't have them. So instead of the AI being able to evaluate what it writes itself, the AI has to try to imitate what humans have written without being able to see if what it wrote is good or not. In chess on the other hand the AI can check itself if it did good moves, since it can play test games and see if the moves leads to victories or not.

If you want to try to code an AI that doesn't understand chess, then write it entirely without encoding any chess rules at all, just train a neural network based on human played games. I'd bet it be hard to even have it do legal moves reliably. That is the level of our current natural language AI.

If you think art is about creating the most realistic depiction of reality, or the most varied representations of an object, or anything else that can be maximised like chess (try to win in as few moves as possible, while minimising the risk of losing), then sure. But then you don't get the first thing about art - you're talking about a particular type of decoration.

It might just be that there is nothing “deeper” to understand about Chess. It’s an artificial game where the rules are the way they are because we said so - i.e. there is no “why”.

I never said AI will not get better. Just that its current level of understanding of many of the fields it’s trying to tackle is similar to that of a cargo cult’s.

The same could be said of music. There's nothing that inherently makes a set of frequencies between 20hz and 20khz played in a particular succession "deep". I don't think AI will ever "understand" it, because most human beings don't understand it. Because there's no inherent value to any of it. Just a confluence rules, personal decisions, circumstances and observations that has lead to the tropes and trends that make up today's - or any given period's - musical landscape.

> There's nothing that inherently makes a set of frequencies between 20hz and 20khz played in a particular succession "deep".

What about the abstract emotional perception of humans that determines if something resonates with them? You can learn that certain chord patterns have certain emotions associated with them. It's another step of difficulty to make an emotionally coherent that makes people feel something. Most art is about individual expression. Not impossible to fake, but in order to do so well you need to understand the context of the world and the salience of certain things to humans. word2vec won't win any awards for its latest single "king - man is a queen"

Chess has a pretty well defined mathematical reward function with perfect information....

Art has never had anything like that

Humans don't fit against exact descriptions.

Humans instead learn based on reinforcement, be it positive or negative. So the human essentially builds a world model of what is "good" and "acceptable", and then works against that.

I suspect that a Reinforcement Learning approach to art generation should be producing art that is much better than current works simply because it doesn't aim to imitate, but instead it aims to maximize rewards, or the fitness of the work without relying on individual measurements against an ideal.

It should be noted that RL research is very costly, perhaps a more feasible approach is to train a model to generate art based on noise, and then train the RL agent to generate good noise.

> The majority of human-produced art is also mechanical, mimicry and repetitive.

Is that so? How much human-produced art have you looked at?

I also feel like a huge piece of art being fun and interesting is the human to human communication aspect of it. Dali, without knowing he’s Dali and what he and the rest of his generation saw, talked about, wrote about, and valued is just a bunch of goofy nonsense, albeit technically well executed.

> How many people can remember their favourite example of fractal art

I’ve got a list of my 10 favorite I can share.

You should actually play with these tools before you dismiss them. There is a lot more going on than bumping around human space. The curation process between artist and machine is live and well in this genre.

I posted a thread [1] on VQGAN+CLIP art generation, curious what you think. I definitely see the points you wrote here, but at the same time I -do- like what it generated. I think there'll be versions that will use patterns better. So they won't understand why we like something, but will be close to picking the correct pattern/features to make us like something.

[1] : https://twitter.com/Killa_ru/status/1439299650331320327?s=19

I think you're right. This is discussed in the article- GPT-3 does very well at the small-scale, like short poems. The architecture of GPT-3 is based on Word2Vec, and attention - it doesn't really have a good understanding of large-scale semantic structures in writing. (Also mentioned in the article - it can't rhyme: it doesn't understand sub-word semantic structure either).

I think the next big advance will be when we are able to apply attention/word2vec-style algorithms at other semantic levels.

AI adds a tool in the toolbox for artists, it will probably never replace human artists.

Reply on news.ycombinator.com