So I think the consensus for using LLMs in creative writing is fairly mixed.
But I was wondering how the devs feel about using it for finding (possible) bugs in the code or creating unit tests etc.
I’m not L&L, but AI has become very widely used in general software development now as in… almost everyone is using it, not just a few people experimenting. I have fewer (!) moral qualms with this given the open nature of coding. I reckon more than 95% of my own code is now via Claude.
I hate this timeline.
2 Likes
Yes, it certainly hasn’t played out as I thought it would.
There are mixed opinions at L&L. I don’t use any AI, although Apple are now pushing agentic coding and other AI options in Xcode. Reading various blog posts and such (e.g. Daring Fireball), it seems I’m a bit of a dinosaur in that regard and that many coders are embracing it. There are others at L&L who do want to embrace it too. Perhaps we should to speed up our notoriously slow development cycles. 
My main qualms are (1) I don’t want our own code to be used to train LLMs, and (2) I don’t want to use any AI that was trained on other people’s code without permission (and, if necessary, payment). And partly because I haven’t had time to look into these things (e.g. how Claude is trained), or, in fact, to investigate how agentic coding works at all, I’ve steered clear. When I have more time, I’ll catch up.
(In general I’m agnostic when it comes to AI as a technology. How it’s trained and how it’s used, on the other hand…)
7 Likes
We are using some agentic workflows at work now: Any member of staff can raise an issue (bug report) in GitHub, and Claude will branch the project and attempt to fix the issue (or even add a feature). It’s… an odd experience. Still requires a skilled engineer to check and merge to code in, but some bug fixes have been almost entirely hands free.
3 Likes
I’m still sitting on the fence, though one cheek is definitely sliding off.
I’ve had some decent results using it for stuff like generating release notes from bug reports, which is great time-saver, as long as you check the output.
But using it to write code? 
I dunno. I like to retain an understanding of the code I’ve written, so I’m not happy to let some AI write it, because I might have to fix it or change it later on — and that’s when the trouble starts.
Then of course, we have this nonsense:
Yeah, well from my experience with Scrivener, you have one of the most robust applications I’ve ever come across, which may have something to do with your “notoriously slow development cycles.”
One great use for LLMs is to automate forum participation:
Craft a forum message that politely suggests that Literature & Latte should do whatever the user inputs as a ventilated prose paragraph. The forum message should be at least 30 lines long and must incorporate the phrase: “leaving money on the table.”
2 Likes