Hacker News Reader

Comment by some_furry | Hacker News Reader

Successful technologies that were hyped up started with acknowledging a problem exists, then working to solve it.

The blockchain / big data / Cloud Everything / LLM hype is largely a solution (the thing being hyped up) looking for a problem.

I hope that's clear.

The problem isn't the hype in and of itself. It's that there's a disproportionate amount of hype for things that aren't actually that valuable. I don't begrudge people for being excited about technology, only for being excited about excitement (and forgetting to actually solve problems for businesses and consumers).

View on HN

If LLMs stop evolving now, we (humanity) have a great new technology to build on. When we have computers and smartphones that can run Llama 3.1 405B the possibilities will be infinite.

If that doesn't turn into something as profitable as Google, that's actually great for all of us but terrible for the investors. I'm actually hoping for this scenario.

I think LLMs have some killer features which solve real problems.

Translation (the problem transformers were invented to solve), Document QA, Embeddings for more semantic search, and unstructured data ingestion are all areas where LLMs excel.

When you have documentation to reference and keep in context, it’s pretty easy to achieve a low hallucination rate.

The hype is around like content creation and AGI, which is a nothingburger.

what is 'document QA' ?

My guess it's answering "how do I do X", where X is on the documentation.

Still I don't understand the people that claim the shared LLMs can do that with internal documentation. That's bullshit and easily disproved, you can force it to work for toy problems, but only for toy problems.

Off the shelf, LLMs can’t just look at docs and give you the answer.

But if you properly pre-process the documents and create a RAG type system (which uses embedding to find semantically similar docs before inserting them into LLM context) then it actually works quite well.

It’s good for big organizations with internal wikis, I’ve found.

It also works well for ingesting articles from online publications.

ChatGPT helped me to setup a hugo page from scratch with templating and stuff.

It did so by just reading the hugo documentation without any RAG.

And it was 10x better than the huge documentation in itself.

Well that’s because Hugo is very simple to set up, there are lots of tutorials online about it, it hasn’t changed drastically in the past year and it’s mostly the same for different static site generators.

You won’t find that kind of ease of use with say webgpu+winit for building a small renderer in rust.

> The hype is around like content creation and AGI, which is a nothingburger.

Right, which is why TFA specifically talks about "Generative AI".

blockchain: selling/buy illegal stuff online

big data: is just data bigger than fix on one computer, many had this problem before and after the term got popular, many more got sold they have the problem without actually having it

cloud: as a user I love being able to get a new workstation running with just installing a password mgr, not loosing files in fires

LLM: being able to search through my pictures by text

Reply on news.ycombinator.com