Rune

@runefar.bsky.social

He/Him/They(gender apathetic cis-male) Half Norwegian American whose pleasures in life include exploring different topics and considering how to build solutions to different issues. Has a background in earth science, anthropology and psychology +tech

I do agree about ai being the wrong people though and thwt is because they are aware of the taste of many people and their own. When we trh to find taste in this sense we are treating taste as if there is just one but there are tons of different cultural ones

It seems like the main difference is the expectation yiu have that this persona would be exclusionary focused and then selecting for new things. Like I admit i think you are under appreciating how much taste is often coming from human boas to nostalgia too

In part i have wondered similar thing myself thinking about people example of hallucinations ans how what they have pointed to as a hallucination comes about in part due to cultural interpretation. I think that would be a interesting aspect when it comes to taste for me is audience interpretation

Though ai has bias, our effort at sorting out bias are ironically often exactilt conflicting it with taste too except when it takes on a specific persona

I should write that up, I guess. But "taste" belongs to a much larger category of constructs where the benchmarking drive that has been so successful in pushing ML forward is going to spin its wheels, I suspect, without deeper reflection. arxiv.org/abs/2411.10939

Evaluating Generative AI Systems is a Social Science Measurement Challenge

Across academia, industry, and government, there is an increasing awareness that the measurement tasks involved in evaluating generative AI (GenAI) systems are especially difficult. We argue that thes...

arxiv.org

or where you went to school. But if we ask, like, why UIs that look obviously Claude-designed are now gross, or why tricolons and antitheses that were "crisp!" in 2018 are now "blech" — that's what I mean. It's a treadmill, not a stable task. Maybe we can train models to run the treadmill? +

It's super fuzzy and complex. a) I'm sure there are some principles of, say, UI design that remain stable, just as some aspects of picture composition are stable. Those would be trainable. b) The relevant kind of status here is status in the field/profession. It's not about how much money you have +

I think this is a bit reductive — what people are referring to when they talk about engineering or design taste is something a little different because people with the same social status have it to very varying degrees

It's possible that tasks of this peculiar kind—tasks inseparable from human status competition—are a rare category only to be found in the fine arts. "Creativity" may be something quite distinct—involving, say, objective diversity & pace of idea generation. It's possible. But I'm not v certain.

Training a generative model to have taste is going to run into the basic problem that generative models are, definitionally, the wrong people. I can see why this will seem frustrating and perverse, if you believe human behavior ought to be rational. But, uh.

What that looks like in practice: take glossy photoreal fantasy art back to DeviantArt c. 2019 and it would be "whoa!" Now? "ugh, slop!" Because the wrong ppl make it. This is not a prejudice we factor out in order to measure true taste. It's how taste works. It's why Monets cost more than LePage.

Jules Bastien Lepage, Rural Love, 1882Monet, Water Lilies, 1903

In general, agree with Ethan here. Certainly sentences that begin "X can never ..." have a bad track record. But with taste ... I actually think discourse around AI may be mistaken in framing "taste" as a task at all. I think we have good evidence that it's about status & group membership. +

Ethan Mollick@emollick.bsky.social · 3h ago

I find arguments that AI can't do judgement or creativity or taste to be especially obviously false in the time of agents. Any long task requires lots of taste, judgement & creativity. I am much more sympathetic to arguments about the quality or diversity of the AI's taste, judgement, or creativity

In my observation of AI chatbots: women are not used to being validated. We also have the bullshit radars of an oppressed class. Sycophancy is annoying but a refreshing change. For men, sycophancy builds the validation of the patriarchy to dizzying heights.

you could severely constrain Palantir’s abilities through strict data privacy and data sharing laws (that also apply to the state). most of how the state Knows Things is by googling it or buying random datasets.

I love that when I was doing organizing in college and grad school I kept preaching opsec to people and no one took it seriously but now they’re all radicalized against idk automatic speeding ticket enforcement, glonzo water usage, and smart glasses rather than idk palantir and dat privacy

I feel like people who previously focused on being anti-surveillance have really lost the plot on ai by pretending that there isn’t demand for it rather than focusing on its role in current & future police states