[leecommamichael]

I'm of the opinion that it can be a tool if used as one, but that most people are currently interested in experimenting with various sci-fi visions. I don't ascribe any emotion or judgment to that, either. We should be doing the things we're excited to do if it doesn't cause too much harm.

There are boring and reliable uses for these things, but then the wins are smaller, so they're not as worth talking about. We all want to say something insightful about the current topic of discussion, and per usual some of the worst behavior gets the most attention.

To add context to what I'm proposing, I think they're good for dealing with issues of scale:

1. search dense files for a precise thing

2. refactor from a "bad way" to a "good way"

3. generate short (<200 line) scripts *

4. getting started with third party SDKs *

5. generate alternative procedures/approaches *

The asterisks denote potentially faulty usage. Short scripts are great to have roughly automated, but as ever the risk with these things has been that they might grow into programs, and that's a poor foundation to build on. This is similar to my rationale with generating example code for unfamiliar SDKs; sometimes usage is not as simple as most guides on the internet, which means you get a sub-par result from the LLM. I think this is pretty much the case with things like win32 or AppKit programming in C. As for the final point, you've pretty much got to be an expert to avoid going through the trouble of entertaining poor suggestions; I find this to be the primary failure-mode of LLMs, they can waste your time.

[jraph]

Your comment focuses on usage. The post argues that we can't say "it's just a tool" because there are inherent traits that don't depend on usage.

> if it doesn't cause too much harm.

You touch something there: some of us think that simply using LLMs at all encourages developing / spreading this tech that inherently causes too much harm.

[leecommamichael]

Yeah, I tried to just tuoch that because it is highly personal. I don't think I entirely understand the post's point about "inherent traits that don't depend on usage." All tools have these traits as well, don't they?

[jraph]

Sure, but making them work doesn't usually boil the planet, exhaust all the resources and risk causing an economical crisis while stealing anything that lies around on the internet.

[zahlman]

To the extent that these claims actually pan out, responsibility should be assigned to the companies developing these systems, not to their product.

[jraph]

Users of these products are also responsible. They are the ones deciding to use them. To the extent, of course, that their knowledge is obviously limited. And to the extent that they are also subject to the competition.

The problem here is that nobody is accountable for the externalities in this economic system where you can do damage as long as you have enough money to throw away toward this damage. Anthropic, xAI, OpenAI and Google obviously don't care. Public powers mostly don't (regulate) neither. If their potential users don't, who is going to?

The sad state of affairs is that nobody relevant cares, as usual.

[zahlman]

> The asterisks denote potentially faulty usage. Short scripts are great to have roughly automated, but as ever the risk with these things has been that they might grow into programs, and that's a poor foundation to build on.

They can be. But you can also refactor (AI assisted or not) before proceeding.

[user43928]

In my opinion AI is good for:

- developing all but the most complex software

- maintaining a >100k LOC codebase on your own

- generating illustrations and animations

My iOS codebase has 95k LOC + another 40k tests, and it is working great. It has been months of work.

[leecommamichael]

What does your application do?

[user43928]

Something like media management.