Hacker News new | past | comments | ask | show | jobs | submit login

Uses for LLMs and Neural Nets in general include: Searching documents in discovery, sentiment analysis, summarization, flushing out ideas, 1st level support, writing code, doing speech->text, text->speech, translation.

There are many things we can do now that were previously expensive or cost prohibitive.

This is about the same as the computerization of records. No longer to we have to pay people to keep rooms full of files in order to aid retrieval and long term record storage.

--- but I see you're asking about non-cost things

We can now read documents that are translated from other languages, lowering those barriers to learning about other cultures. It's now easier to interact with the information available to all via the internet.

It also makes some evil things easier, but that comes with every technology.




OK, I can partially agree with some of the points but... writing code? Any tangible proof on that one? I have spoken with many colleagues who were adamant that the LLMs can at best provide some boilerplate generation and nothing much beyond that. One or two mentioned it could find a bug in their Python code.

> This is about the same as the computerization of records. No longer to we have to pay people to keep rooms full of files in order to aid retrieval and long term record storage.

Eh, have you read the "Bullshit Jobs" book? It's a social and economic problem, rarely has been a technological one... TL;DR people want headcount, not efficiency. :/


> writing code? Any tangible proof on that one?

Personal experience? Not just me, but e.g. all the comments on HN on ChatGPT, Copilot, et al.?

I’ve been using it for about two years now to draft new code, after having programmed for two decades. It provides value to me in that some tasks that previously would take me a whole day now can be done in a couple of hours, maybe one hour of throwing ball with ChatGPT and then one hour of refactoring the code myself. I have plenty of colleagues who have similar experiences, especially when using it for tasks that can be “tricky to write but easy to verify” (e.g. visualizations).

Just keep in mind that (i) the paid version is miles ahead of the free version for coding, and (ii) using an AI assistant is like any other skill, it requires some training to learn what tasks they’re good at and how to talk to them.


Saving programmer time writing code does not, in and of itself, "objectively improve society" as the OP above sought examples of. Perhaps, if you are doing something extraordinary. However, AI also makes spambots and malware easier to write, as well as enormous amounts of simply mundane stuff - shopping cart APIs or credit card comparison websites and ad servers.


Fair enough, but I feel that the same can be said for e.g. the combustion engine…

The only thing cars did was to save time and resources doing things humans were already capable of doing. Now, such engines are used for life saving things like ambulances and food transport, but also nefarious things like invading other countries, or mundane things like dropping off your kids at karate class.

I suppose whether you think the combustion engine actually improved society depends on who you ask as well :)


Huh, didn't realize humans could fly or go to the moon without combustion engines.


I guess we’re stretching the analogy a bit far… But we’re mostly flying using jet engines, which is a distant descendant of a combustion engine that was developed over a century later than the car engine. If we want to see the AI analogue of this, it’s probably a future AGI and not what we have today :)


"We" aren't stretching any anology. Your statement was ridiculous from the start. Pat yourself on the back for saving a couple hours at work, and leave it at that.


> I have spoken with many colleagues who were adamant that the LLMs can at best provide some boilerplate generation and nothing much beyond that.

Why don't you try it out yourself, eh? there are a lot of free options


Because my time is valuable, dude, and having to use an entire weekend -- likely much more than that even -- is not very feasible. And because I trust my circle of programmer acquaintances, they are extremely pragmatic folk who usually gets called to fix messes left by juniors, wannabe, poseurs and "AI" enthusiasts.

Still, I am not entirely close-minded on "AI" but I don't think it's unfair to say that if it was that good then our area would have been hugely disrupted by now. And it does not seem like it is.




The deadline for YC's W25 batch is 8pm PT tonight. Go for it!

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: