[gmerc]

I see some potential for an LLM to unfuck the dom - ultimately there’s going to be patterns

[skulk]

feeding an entire DOM tree into an LLM and waiting for the response sounds like a terrible thing to include in a latency-sensitive render pipeline.

[gmerc]

Looks like apple thinks so too

https://www.macrumors.com/2024/04/30/safari-ai-features-comi...

[gmerc]

longer than loading dozens of shit js, tracking pixels and obfuscation code into the renderer?

The moment I can use a cheap edge LLM to reduce the entire page to the few bullet points I care about with optional expansion of content I don’t really care much anymore about render timing - the ultimate cost on us is extraction the information we want from the page, which is getting more and more enshittified because our attention is being sold.

[dkraj]

I see your point. Getting the pattern from DOM to create something custom is possible but not in real time. We can not just rely on that. Finding pattern is good but relying on that is not