Ask HN: Why does HN think AI image models will never be satisfactory? | Hacker News Reader

Ask HN: Why does HN think AI image models will never be satisfactory?

2 pointsgametorch11 comments

I recently did a "Show HN" for a project I'm building that uses image models and I received great, positive feedback from most people, especially people in real life, and yet tremendous pushback from a vocal minority on Hacker News.

Every one of the comments pushing back implied that AI image models will never be sufficient for the poster's standards.

In the two months since I began working on this project, model quality has increased by an order of magnitude while costs have done the opposite.

I was also able to use LLMs to launch a full featured, production-grade software service in two months that survived the Hacker News "hug of death" without so much as the blink of an eye.

Why is a significant subset of HN so confident when it comes to their view that the exponential improvement curve will not apply to this particular technology? Isn't it folly to bet against the advancement of technology?

This is especially confusing to me when hundreds of billions of dollars and PhDs and professors are thrown at the problem, which has a clear financial incentives aligned with finding the best solution. (Obligatory "this isn't nuclear fusion!")

Only one argument has made sense to me: AI lowers the bar for releasing stuff into the wild. This means you'll see more things and those things will be, on average, worse in quality than what you saw before. The argument leads on to say this pent up, subconscious distaste for AI-related crap is what leads to pushback. Fair enough.

The rest of the arguments that make sense to me follow a similar structure but are fraught with logical fallacy --- "AI is replacing jobs" or "AI is destroying the earth" are very interesting topics that should be investigated, revisited and reviewed periodically, but ultimately these claims speak against the idea of allowing AI to be developed and used; they say nothing of its quality.

AI models have added tremendous value to my life already. I've been glad to pay for it all. We are on a clear "up and to the right" trajectory in terms of quality. What gives? Why does a significant subset of Hacker News think quality is not going to go up and to the right?

HN these days is very similar to Reddit. Most users spend their free time talking about things others have done, rather than doing things themselves. Of course, most of this internet discussion leans negative for various reasons that have been addressed in other places better than I could recap.

Why care what anybody on here thinks? They're mostly anonymous nobodies.

Don't worry about HN, they represent 0.000000001% of the entire human population, it is essentially a very very small bubble.

As for why does HN think AI image models will never be satisfactory?

They don't have to be satisfactory, it just has to be good enough for the vast majority of people.

Remember what everyone said about film and digital? Give it time. Take every advantage open to you. Satisfy someone, if not this mob.

> Isn't it folly to bet against the advancement of technology?

I just like good art. I don't have any strong feelings for or against AI, but I do epistemically reject art that lacks composition or intent. AI-generated art doesn't understand rhythm, symbolism or image arrangement. This is an obvious problem when trying to generate a photoreal subject without six fingers per hand, but especially troubling if you want to give Starry Night a run for it's money.

You will be perpetually disappointed if you portray this as a "luddite vs enlightenment" problem. If you enjoy AI art, more power to you! The rest of us are overwhelmingly disinterested, AI art isn't filling out any gallery I've ever visited.

> Why is a significant subset of HN so confident when it comes to their view that the exponential improvement curve

Which scaling law has ever promised an "exponential improvement curve" for image generation, let alone AI as a whole? I think you're making stuff up here, or we're reading different research papers.

I definitely see what you're with the art part of this.

My comment here summarizes my feelings: https://news.ycombinator.com/item?id=44207935

Respectfully, I still disagree. Art is not commoditized, even if you can appear to use a replacement product. People called out the issues with your pixel art immediately when you uploaded it, clearly AI art isn't actually an art commodity. And of course, look at the world of "vibe coding" - it hasn't even remotely commoditized software development. It's really hard to write good code, and the same goes for making good art.

Again, I sympathize with your pragmatism here. Having AI-generated placeholder assets beats programmer art, but you're just blatantly wrong if you expect "exponential improvement" apropos of your own observations.

The answer to your original question is likely that HN perceives AI capabilities to be a bathtub curve and not an exponential climb.

It has though.

I used commoditized LLMs to build an entire website with paying users in 2 months.

Cost me <$200.

It stood up to the Hacker News hug of death without breaking a sweat.

The generation of code and pixels is being commoditized, whether you like it or not.

Asking for opinions then telling people their opinions are wrong is silly and will discourage anyone else who might be willing to have a good faith convo about your original question.

Am I supposed to lie?

I think generative AI is commoditized. The commenter denied that but gave an explanation about something else.

Whether generative AI constitutes art is a separate question merely of semantics, about which I did not deny the commenter their own opinion.

You're the one acting bad faith, it feels. I spoke truth about commoditization. I did not deny them their opinion about the semantics of what constitutes art.

I can explain it to you, I just expect a good-faith response in return. From the outside looking in, it doesn't even feel like you're trying to argue your point. Your example is a self-made webapp, which seems to be unsuccessful and not very popular among game developers. You have one success story, your own, and it isn't for a popular video game or a difficult, creative task.

How am I supposed to believe you when other people don't want what you have?

From your other comment:

> I can generate a really good looking one for 10 cents because this component of art has been commodified.

You can 3D print a decent carburetor, but that doesn't mean that component of car manufacturing has been commodified. You have to compete in the market, and AI-generated art (much like 3D printed carburetors) has to live in it's own niche submarket. It is not an acceptable substitute for downstream consumers of this medium. Look at the backlash to Marvel's Secret Invasion when they thought AI was "good enough" for prestige television.

Nobody's going to stop you from chasing this dream, but I'll tell you outright that I don't want to pay for a game with AI art. The games industry is commodified well enough that I don't have to settle for your product.

> which seems to be unsuccessful and not very popular among game developers

It just launched and we already have paying users.

Reply on news.ycombinator.com