this post was submitted on 03 Jun 2024
1474 points (97.9% liked)
People Twitter
5234 readers
838 users here now
People tweeting stuff. We allow tweets from anyone.
RULES:
- Mark NSFW content.
- No doxxing people.
- Must be a tweet or similar
- No bullying or international politcs
- Be excellent to each other.
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Indeed. LLMs read with the same sort of comprehension that humans have, so if a supermarket makes their website compatible with humans then it's also compatible with LLMs. We have the same "API", as it were.
Can LLMs interpret structured input like html?
Yup. And those that can't can have a parser pull just the human-readable text out, like a blind person's screen-reader would do.
That sounds like an issue with your system prompt. If you're using an LLM to interpret web pages for price information then you'd want to include instructions about what to do if the information simply isn't in the web page to begin with. If you don't tell the AI what to do under those circumstances you can't expect any specific behaviour because it wouldn't know what it's supposed to do.
I suspect from this comment that you haven't actually worked with LLMs much, and are just going off the general "lol they hallucinate" perception they have right now? I've worked with LLMs a fair bit and they very rarely have trouble interpreting what's in their provided context (as would be the case here with web page content). Hallucinations come from relying on their own "trained" information, which they recall imperfectly and often gets a bit jumbled. To continue using a human analogy, it's like asking someone to rely on their own memory rather than reading information from a piece of paper.
Or you could just prompt it to not guess prices for articles that don't exist. Those models are pretty good at following instructions.