Comments URL: https://news.ycombinator.com/item?id=48248284

Points: 4

# Comments: 1

Comments URL: https://news.ycombinator.com/item?id=48167116

Points: 4

# Comments: 0

If there's any hope for reliability, auditability, predictability to be had it lies in contraining and LlMs grammar whilst delegating freeform behavior to a more passive substrate.

The problem with products like this is that for search to be useful, you need, well... data. And when your data is crowdsourced, you need, well... a crowd. That was the chicken-and-egg situation I found myself in.

So the way I solved it was by painstakingly clipping and recording ~8500 websites myself.

But the great thing about that arduous process is that it forces you to put yourself in the shoes of your typical user. And really, the only way to make good products is to have a deep level of empathy for that user. In my case, that meant streamlining the data creation process itself, the act of pinning a screenshot or recording, because I had to do it over and over again. Making it easier for myself ended up making it better for everyone else too.

Of course, all that data-seeding effort goes to waste if search engines and potential users cannot discover it. So you have to go all in on SEO, especially sitemaps, and on programmatic socials where possible. This is especially true for a product like mine.

A lot of these marketplaces also have relatively homogenous data, which is actually useful, because it means you can present it through templates and publish pages at scale automatically.

Agents can search for design inspiration from production websites using semantic search. Since this inspiration comes from live websites, their design tokens; colors, typography usage, layout data are also available.

I understand that finding an excerpt nearly half a century old, from a little-known magazine and a little-known author, is bound to be challenging.

Comments URL: https://news.ycombinator.com/item?id=47135571

Points: 3

# Comments: 0

It becomes a question of taste. LLMs create exceptionally generic and inconsistent designs in the absence of curated references.

FontofWeb grounds AI in real production websites. It draws from a crowd-sourced search engine of over 1,000 live sites, extracting proven color and typography combinations, along with layout and styling patterns that already work in context.

This is available through a free public MCP service.

Comments URL: https://news.ycombinator.com/item?id=47041394

Points: 1

# Comments: 0

Would instead of the RL step a constrained decoding say via something like xgrammar fix syntax generation issue ?