Hacker News Reader

Comment by b112 | Hacker News Reader

So far, I haven't seen a useful application of LLMs. So far.

I've seen things that are wildly hobbled, and wildly inaccurate. I've seen endless companies running around, trying to improve on things. I've seen people looking in wonder at LLMs making mistakes 2 year olds don't.

Most LLM usage seems to be in two categories. Replace people's jobs with wildly inaccurate and massively broken output, or trick people into doing things.

I'd have to say Bitcoin is far more useful than LLMs. You have to add the pluses, and subtract the minuses, and in that view, LLMs are -1 billion, and bitcoin is maybe a 1 or 2.

View on HN

AI is not just LLMs. AlphaFold for example moved a critical goal post for everyone of us.

bitcoin is only negative. It consumes terrawatts of energy for nothing.

And even if it were just LLMs, I use LLMs in my workflow every single day, and I've never used a/the blockchain except for some mild speculation around 2017.

I'm as skeptical about LLMs as anyone, especially when people use them for actual precision tasks (like coding), but what they actually IMHO are good at are language tasks. That is, summarising content, text generation for sufficiently formulaic tasks, even translation to an extent, and similar things.

There is one clear (albeit somewhat boring) application of LLM: data extraction from structured documents.

That field has made a leap forward with LLMs.

Positive impact on society includes automated extraction in healthcare pipelines.

Unstructured*

No, I really meant structured. Extracting data from structured documents is surprisingly hard when you need very high accuracy.

What I mean by structured is: invoices, documents containing tables, etc.

Extracting useful data from fully unstructured content is very hard IMO and potentially above the capacity of LLMs (depending on your definition of "useful" and "unstructured")

But this is why I made my complexity statement in my other reply.

Why are firms sending around invoices, tables instead of parseable data. Oh I know the argument, because "so hard to cooperate" on standards, etc.

Madness.

Partly because the standards, such as X12, have a high startup cost to use them, they aren't very opinionated about the actual content, and you have to get the counterparty on board to use them.

Healthcare pipelines! All well and good until hallucinations cause death or what not!

And why is this better than employing a human. Or reducing complexity. It's not as if human wages are what causes hyper expensive US healthcare costs.

This seems like a negative.

Right now there is no human, the data just goes nowhere (i.e. it is not used).

At some point we need to be optimistic and look for incremental progress.

> So far, I haven't seen a useful application of LLMs. So far.

What?! Whole industries have been changed already due to products based on them. I don't think there's a single developer who is not using AI to get help while coding, and if you aren't, sorry but you're just missing out, it's not perfect but it doesn't need to be. It just needs to be better than StackOverflow and googling around for the docs or how to do things and ending up in dubious sites, and it absolutely is.

My wife is a researcher and has to read LOTS of papers. Letting AI summarize it has made her enormously more efficient at filtering out what she needs to go into more detail.

Generating relevant images for blog posts is now so easy to do (you may not like it, but as an author who used to use irrelevant photos before instead, I love it when you use it tastefully).

Seriously, I can't even believe someone in 2024 can say there has not been useful applications of LLMs (almost all AI now is based on LLMs as far as I know) with a straight face.

> I don't think there's a single developer who is not using AI to get help while coding

You are in a bubble.

> It just needs to be better than StackOverflow and googling around for the docs or how to do things and ending up in dubious sites, and it absolutely is.

Subjectively. Not absolutely.

> I don't think there's a single developer who is not using AI to get help while coding

It's banned at my company due to copyright concerns. Company policy at the moment considers it a copyright landmine. It does need to be "perfect" at not being a legal liability at the very least.

And the blog post image thing is not a great point. AI images for blog posts, on the whole, are still quite terrible and immediately recognizable as AI generated slop. I usually click out of articles immediately when I see an AI image at the top, because I expect the rest of the article to be in line: low value, high fluff.

There are useful LLM applications, but for things that play to its strengths. It's effectively a search engine. Using it for search and summarization is useful. Using it to generate code based on code it has read would be useful if it weren't for the copyright liability, and I would argue that if you have that much boilerplate, the answer is better abstractions, libraries, and frameworks, rather than just generating that code stochastically. Imagine if the answer to assembly language being verbose was to just generate all of it rather than creating compiled programming languages.

Reply on news.ycombinator.com