Hacker News Reader

Comment by jsheard | Hacker News Reader

The same goes for image generation models, AI art already has a tendency to veer into the same clichés and those are only going to get reinforced if newer models are trained on newer scrapes which now include the million hyper-derivative AI images being uploaded to places like DeviantArt, Twitter and Pixiv every day. Those vendors who got in early have a moat in the form of untainted scrapes, but they'll eventually need new data to keep up with new subjects/styles/etc.

View on HN

Art is much less vulnerable.

Art is very heavily tagged, accurately described and filtered. Many galleries don't accept AI art at all, which means to sneak in the AI needs to be pretty much perfect.

Also, there's an enthusiastic scene of LoRAs where the makers work with small enough datasets to do manual curation.

Photography has been very well developed technically for almost a century. Yet very little is considered art. The most interesting AI art, to me, are “image compilations” where the compiler appears to have taken just a bit of LSD.

>> which means to sneak in the AI needs to be pretty much perfect

That's status quo. Because right now artificial art is easy to distinguish form "natural art". Besides that I consider this - no offense! - as some kind of "arrogance". Galleries accept what they think what art is. Why can't I, the art consumer, decide for myself, what art is?

Are 'scrapes' really the only acceptable source?

Not always, a few companies like Getty Images are using their internal stock photo libraries to train models rather than slurping up the entire internet. They are uniquely positioned to take that path though, having already been in the stock photo business for decades before AI took off, and for everyone else it's a sisyphean task to manually curate a complete representative dataset that they can be sure isn't contaminated with AI images.

This is such a wrong take.

Even if the data is 100% synthetic, you can still hill climb to new mountains.

If you don't believe me, look at evolution.

It doesn't matter if we no longer have 100% human art as input. This is the worst these systems will ever look and feel, and they're only going to improve.

I'd be willing to do a longbets on this one.

How does evolution use synthetic data?

Hill-climbing requires some measure to optimize. current technology uses 'how well can you predict real text' as this measure. If you change the text you try to predict, it changes the measure you are optimizing. It is far from obvious that this will still improve the actual performance.

Evolution has a clearly defined fitness function (number of offspring that are produced). What is the fitness function for generated art?

Exactly, the fitness function is "make a picture which looks like these other pictures", which is used as a proxy for "make a picture that a human would find appealing by making it like all these pictures made by humans". The former definiton will always hold true, but the latter actually desired definition breaks down once your training set is contaminated with imitations of imitations of imitations.

Number of works that are released to the internet! At the end of the day, there is a human who has an idea in mind and is using an LLM to realize it. They are tweaking their prompt until they achieve their vision, then publishing only those successes. The resulting art can then be used with the prompt as training data.

It would be nice if every human at the wheel of an LLM or image generator were putting that much care into their craft, but that's just not what's happening. You just need to look at places like DeviantArt which are now overrun with users who joined 6 months ago and already have submissions in the thousands, it's simply not possible that they are putting any care into what they're spraying out. Much of the time they're posting numerous functionality identical images, probably generated from the same prompt with a different seed.

Likewise SEO incentivizes mass production of low quality LLM text, because they "quality" they are optimizing for is impressions, not actual quality.

That is assuming there is a human in the loop. What if an SEO company is using a bot to generate vast amounts of images with no human in the loop?

It's getting really tiresome to see tech bros talk about biology as if they know what they're talking about.

> It's getting really tiresome to see tech bros talk about biology as if they know what they're talking about.

My undergrad was in biology, "bro". I was cloning luciferace into plants using agrobacterium-mediated transfection over a decade before this week's news of transgenic petunias.

I was planning to do a PhD in computational metabolomics, but life took a different turn: a couple of Google engineers saw the laser projector I built and programmed to play video games on the side of skyscrapers [1], and they lured me away to work on what became a decacorn. My quality of life and net worth certainly thank me for the pivot.

Take a look at my post history [2]. My analogies are informed and salient.

In any case, I'm working directly in this field now and the results we're achieving don't need affirmation. I tried to communicate this to outsiders in an easy to digest analogy. I'll let our work speak for itself.

[1] https://youtu.be/5XTi-jf-ans

[2] https://hn.algolia.com/?dateRange=all&page=0&prefix=false&qu...

If you studied biology then you ought to know better.

You need to study optimization. I made a perfectly salient analogy. I'm incredibly busy working, and I'm not going to write an essay to bridge the gap for you. All of the pieces are there in front of you. Humans and the free market are an exogenous fitness landscape that is more than suitable to push these systems to their pinnacle.

perfectly, no chance at all for imperfection. entirely sure of himself, he is.

I'm completely fallible. I just don't entertain people calling me a "tech bro" that lazily follow up with "but you should know better". That's not earnest, good faith engagement.

I assure you I'm being entirely earnest here.

There are biological analogies you could have chosen other than adaptive hill-climbing. For instance, there's adaptive radiation, where a small population of organisms is introduced into a new environment and rapidly diversifies into new species, filling niches that weren't filled before. After the environment is mostly saturated, the populations stabilize and adaptation slows down, with gradual and relatively slow change and diversification.

Plugging LLM into this analogy would lead to a story where we're at the initial "Permian explosion" phase of evolution. New LLMs are rapidly "hill climbing" as they feed on virgin data scraped from the Internet. But as those LLMs diversify and adapt, they peter out once all niches have been filled and the fuel consumed.

Another analogy is the Petri dish, but that's an even more pessimistic analogy than adaptive radiation.

> After the environment is mostly saturated, the populations stabilize and adaptation slows down, with gradual and relatively slow change and diversification.

We have the pressures of growth and novelty. My analogy is perfect.

Reply on news.ycombinator.com