Hacker News Reader

Comment by bulbosaur123 | Hacker News Reader

Disagree. Just allow one or two more iterations and it will supersede human abilities. Think ahead. Tech progress won't stop.

View on HN

The tech will get better, but ultimately there still has to be a human who decides 'that's the one that looks good', which strongly depends on someone's taste and skill in identifying what a good image looks like.

There will probably be less need for designers of 'lower quality' simple images though.

I agree with you, but what if what constitutes good taste is just a subset of things that we’ve seen and liked.

If dale decides what we see, it might become what the next generation likes and considers “good taste”.

This is an interesting conversation. Good taste is what we see and like … but also patterning after people we want to impress / be associated with, is it not?

Taste is very complex: it's hierarchical, social, not fixed, not absolute, not rational, is specific to audience and has irregular overlaps across groups, much of it (all?) derived from human sensation and context-specific situations.

The path to something being considered as good taste is generally not simple: much of it flows through lines of power/desire/moment whose branches are not easy to trace as they're being formed. Much of taste is the hidden "why" which most of us never see.

It's realistic that Dall-E could understand what trends are on the rise, or in good taste … it's much harder to say if Dall-E could create something of originally good taste.

That just sounds like pattern recognition with extra variables. Subdividing people into groups and then analyzing them certainly doesn't sound like a task that a machine will struggle with. Why should the algorithm need to be able to see the hidden "why" when most of us creative types can't see it or define it either? It's just a function of having observed enough people of a certain type. You want to generate something that will impress the people I'm targeting? Just analyze the posts of all my followers on social media. Analyze the content that is "liked" by people in my demographic range and with close proximity to where I live. Analyze the works of creators who belong to my generation and who listen to the same music as me. Do that all nearly instantly and then offer me a selection of options picked from those various methods. I don't expect "good taste" will be hard to conjure up. I already can't tell that a lot of these octopus drawings weren't created by a talented human, and we're still early and unsophisticated in our data analytics.

> has to be a human who decides 'that's the one that looks good'

Assuming the status quo, true. As we evolve our lives around emerging AI tech I think we will at first be the curators and creative directors of AI, but eventually a creative agency will defer to the AI as it knows more about our tastes, market, audience, and the ENTIRE HISTORY of art, design, marketing, tastes, trends, and so on.

Eventually it won't make sense to have a stupid human rubber stamp what the all powerful AI suggests. Just as it does not make sense for Facebook to curate news feeds.

Maybe one day product advertising will look different depending on who looks at it. Pepsi logo "just for you".

Is anyone really happy with AI curated feeds? Besides the company's who make them?

I am! TikTok is amazing and the ads I get on Facebook/IG are for things I often want to buy.

There has to be x - y humans that needs X - Y hours instead of X humans needing X hours. And that is a real risk to the profession

Only if you assume the world demand for raster images is fixed...

I still remember a HN article, might have been a Paul Graham article, from 15 years ago about “Why are all of Trump’s buildings so poorly designed when he can afford the best designers?” It came down the the fact the he personally has bad taste and therefore cannot pick good designers or approve good designs.

That aside, a great use of these tools is to generate N spit-takes of wildly varying styles that you can present to the customer very quickly and very cheaply. Once you pin them down to a particular range of styles you can get down to the carving out the details by hand.

What looks good is a more widely distributed skill. A lot of people can tell you what looks good quite well, very few people can make it.

However, the input might stop.

Right now, the input to DALL-E is all human generated.

What will happen is that DALL-E will generate something "close enough" that gets used and promulgated, so now the input to DALL-E will become increasingly contaminated with output from DALL-E.

We're already starting to see this in search engines where you get clickbait that seems to be GPT-3 generated.

Reply on news.ycombinator.com