I do read it. In my experience the project will quickly turn into crap if you don't. You do need to steer it at a level of granularity that's appropriate for the problem.

Also, as I said, I've been coding for a long time. The ability to read the code relatively quickly is important, and this won't work for early novices.

The time saving comes almost entirely from having to type less, having to Google around for documentation or examples less, and not having to do long debugging sessions to find brainfart-type errors.

I could imagine that there's a subset of ultra experienced coders who have basically memorized nearly all relevant docs and who don't brainfart anymore... For them this would indeed be useless.

▲

noduerme 4 days ago | parent [-]

I mean, I'm curious what kind of code it's saving you time on. For me, it's worse than useless, because no prompt I could write would really account for the downwind effects in systems that have (1) multiple databases with custom schema, (2) a back-end layer doing user validations while dispatching data, (3) front-end visual effects / art / animation that the LLM can't see or interpret, all working in harmony. Those may be in 4 different languages, but the LLM really just can't get a handle on what's going on well enough. Just ends up hitting its head on a wall or writing mostly garbage.

I have not memorized all the docs to JS, TS, PHP, Python, SCSS, C++, and flavors of SQL. I have an intuition about what question I need to ask, if I can't figure something out on my own, and occasionally an LLM will surface the answer to that faster than I can find it elsewhere... but they are nowhere near being able to write code that you could confidently deploy in a professional environment.

▲

maccard 4 days ago | parent | next [-]

I’m far more in the camp of not AI than pro LLM but I gave Claude the HTML of our jira ticket and told it we had a Jenkins pipeline that we wanted to update specific fields on the ticket of using python. Claude correctly figured out how we were calling python scripts from Jenkins, grabbed a library and one shorted the solution in about 45 seconds. I then asked it to add a post pipeline to do something else which it did, and managed to get it perfectly right.

It was probably 2-3 hours work of screwing around figuring out issue fields, python libraries, etc that was low priority for my team but causing issues on another team who were struggling with some missing information. We never would have actually tasked this out, written a ticket for it, and prioritised it in normal development, but this way it just got done.

I’ve had this experience about 20 times this year for various “little” things that are attention sinks but not hard work - that’s actually quite valuable to us

▲

shsusha 4 days ago | parent | next [-]

> It was probably 2-3 hours work of screwing around figuring out issue fields

How do you know AI did the right thing then? Why would this take you 2-3 hours? If you’re using AI to speed up your understanding that makes sense - I do that all the time and find it enormously useful.

But it sounds like you’re letting AI do the thinking and just checking the final result. This is fine for throwaway work, but if you have to put your name behind it that’s pretty risky, since you don’t actually understand why AI did what it did.

	▲	maccard 4 days ago \| parent \| next [-]
		> How do you know AI did the right thing then? Because I tested it, and I read the code. It was only like 40 lines of python. > Why would this take you 2-3 hours? It's multiple systems that I am a _user_ of, not a professional developer of. I know how to use Jira, I'm not able to offhand tell you how to update specific fields using python - and then repeat for Jenkins, perforce, slack. Getting credentials in (Claude saw how the credentials were being read in other scripts and mirrored that) is another thing. > This is fine for throwaway work, but if you have to put your name behind it that’s pretty risky, since you don’t actually understand why AI did what it did. As I said above, it's 30 lines of code. I did put my name beind it, it's been running on our codebase on every single checkin for 6 months, and has failed 0 times in that time (we have a separate report that we check in a weekly meeting for issues that were being missed by this process). Again, this isn't some massive complicated system - it's just glueing together 3/4 APIs in a tiny script in 1/10 of the time that it took me to do it. Worst case scenario is it does exactly what it did before - nothing.
	▲	noduerme 4 days ago \| parent \| prev [-]
		Hah, even the concept of putting your name behind something is so great. It's kind of the ultimate protest against LLMs and social media, isn't it?

▲

noduerme 4 days ago | parent | prev [-]

I've used it for minor shit like that, but then I go back and look at the code it wrote with all its stupid meandering comments and I realize half the code is like this:

const somecolor='#ff2222'; /* oh wait, the user asked for it to be yellow. Let's change the code below to increase the green and red /

/ hold on, I made somecolor a const. I should either rewrite it as a var or wait, even better maybe a scoped variable! /

hah. Sorry I'm just making this shit up, but okay. I don't hire coders because I just write it myself. If I did, I would assign them all* kinds of annoying small projects. But how the fuck would I deal with it if they were this bad?

If it did save me time, would I want that going into my codebase?

	▲	maccard 4 days ago \| parent [-]
		I've not found it to be bad for smaller things, but I've found once you start iterating on it quickly devolves into absolute nonsense like what you talked about. > If it did save me time, would I want that going into my codebase? Depends - and that's the judgement call. I've managed outsourcers in the pre-LLM days who if you leave them unattended will spew out unimaginable amounts of pure and utter garbage that is just as bad as looping an AI agent with "that's great, please make it more verbose and add more design patterns". I don't use it for anything that I don't want to, but for so many things that just require you to write some code that is just getting in the way of solving the problem you want to solve it's been a boon for me.

▲

svara 4 days ago | parent | prev [-]

I've also not had great experiences with giving it tasks that involve understanding how multiple pieces of a medium-large existing code base work together.

If that's most of what you do, I can see how you'd not be that impressed.

I'd say though that even in such an environment, you'll probably still be able to extract tasks that are relatively self contained, to use the LLM as a search engine ("where is the code that does X") or to have it assist with writing tests and docs.

	▲	jason_oster 3 days ago \| parent [-]
		Your conclusion is spot on. Fuzz generators excel at fuzzy tasks. "Convert the comments in this DOCX file into a markdown table" was an example task that came up with a colleague of mine yesterday. And with that table as a baseline, they wrote a tool to automate the task. It's a perfect example of a tool that isn't fun to write and it isn't a fun problem to solve, but it has an important business function (in the domain of contract negotiation). I am under the impression that the people you are arguing with see themselves as artisans who meticulously control every bit of minutiae for the good of the business. When a manager does that, it's pessimistically called micromanagement. But when a programmer does that, it's craftsmanship worthy of great praise.