I've finally popped my own AI assisted coding (rather than debugging only) cherry. I now have an extremely dim-witted intern, who is overly agreeable with anything that I say, that I can leave to work for 5 minutes while I do something else, then prod along to do the next sub-task.
greenfield or brown?
Mixed? Existing codebase with 100k LOC, building extensions to existing features, implementing bug fixes, etc.
it's probably irritating for people who know how to code but it's great for someone like me who needs a simple task and has no interest in learning. I need to search multiple SRT files for a text string. In 30 seconds I have a script that outputs a txt with file names, line numbers, and strings.
IDGAF about learning python. I'm at the level where I need to google "how to run python script." Chat GPT holds my hand. I do this kind of thing once a month so it's not worth the hours I'd need to spend learning to do it myself.
I have not had the greatest experience with it, although perhaps that’s because my (recently former) employer had a lot of trouble finding a code assistant vendor that could comply with their security requirements. But I found it awesome for some very specific use cases, like
writing basic code against some big, complicated, commonly used API that I have never used before. For instance, I wrote some bookmarklets that extract the title from the current github/JIRA/etc. page and put it on the clipboard as a link ready to pass into a document or chat.
I don’t do front end programming for a living. Though I know the basics of Javascript, browser APIs are Greek to me, and I have no experience debugging it all. But there are lots of examples of those APIs in use all over LLM training sets. And sure enough, LLM based code generation worked great.
That's.. actually probably better than the ai tool we have at my work.
I treat it like pouring concrete. It's a hell of a lot less work than laying bricks, but there's now a ton of preparation you have to do to make it work properly. You can stub out a bunch of classes/function sigs/schemas in a nice testable structure, then have it implement them and write tests.
Like concrete you also can't trust it to figure out things on its own. If the concrete gets outside your spec, its going to create a huge mess thats more work to clean up than just doing it right in the first place.
I’ve come to describe the little LLM thing I’m working with as not even an intern but a college student two years into CS consumed by dunning-Kruger ready to complete any task provided with absolute conviction
I wish I could be optimistic. I give it a try maybe once a week, and every time it fucks up something critical and I have to revert big chunks of it. Even basic refactoring tasks like “rename this C++ class” frequently fail I’ve had _way_ better interns, ngl
TBH, I've never trusted this stuff enough to try anything more than code complete.
Actually, continuing on this thought, my hangup is that it feels like none of this beyond code completion should work. The fact that it clearly works *at all*, even just some of the time, is astonishing, but that fact really doesn't fill me with confidence.
are you using Claude Opus 4.1? that seems to be the trick for us. the 3.x models are just _dumb_.
Internal-only Gemini variant.
Ah! My condolences.
no wonder middle management and CEOs think it’s revolutionary
I can both participate in a Q&A session and also write code at the same time, it's amazing.
Have fun finding all the subtle errors. My guess is that you're going to end up spending more time fixing things than was saved in 'delegation'.
This is why we have unit tests and self-code-review (of generated output, but also just good practice to read your own diffs before sending for review) and peer code review.
Which is great, but unit tests can only catch regressions. And if vetting the AI's output adds much more time in the stages of review, it hasn't helped. It's also going to poke holes in a layer of the swiss cheese model. And the more people start using AI for tasks, more layers get more holes.
Almost as advanced as an F1 monkey in fleet. 🌽😹🌽
what are you using? I need to try for myself too, I feel like it’s time.
Claude Code. @honeycomb.io gave everyone subscriptions.
Cool, thanks. Seems like the popular choice right now.