Remix.run Logo
▲ MattRix 2 hours ago

I don't see any reason all the things you mentioned cannot be taught to an LLM using reinforcement learning. The newer models are already much better at this stuff, I see no reason that trajectory will change.

▲rickdeckard 2 hours ago | parent [-]

Hence I wrote: "Structured data of actual usage would be needed for AI to ingest, in a way that it can derive intent from every action and identify friction."

The challenge is more about providing actual, objective data to learn this.

The LLM will not be able to decide what is "good UX" and "bad UX", it will have to ingest this information in vast detail from judgement provided by humans, which is also varying alot.

Case in point: Coming from that field, I personally don't have a good user experience using iOS, a majority of end-users will strongly disagree.

An LLM would need to be trained to handle both cases to judge the same human interface, so it would need alot of data to understand both views...

▲adrianN an hour ago | parent [-]

The human UI designer has the benefit of being and to run experiments on themselves, but only n=1. for any problem where that is insufficient or even misleading that benefit disappears. I’d argue that UI is an area where you need n>>1.

▲rickdeckard 34 minutes ago | parent [-]

That's what UX user studies are for, which is a process of humans observing other humans using the product and taking notes. It's far more common than people seem to believe.

(also: UX is not UI, but I guess this goes without saying...)