Ask HN: Do you use AI to generate majority of your production code?
These days I often see extraordinary claims from reputable industry and other insiders that AI writes most of the code and human programmers are simply supervising it.
For example, this morning I was reading [1] and it had the following claim: “ If you really want to grasp how much better A.I. has gotten recently, talk to a programmer. A year or two ago, A.I. coding tools existed, but were aimed more at speeding up human coders than at replacing them. Today, software engineers tell me that A.I. does most of the actual coding for them, and that they increasingly feel that their job is to supervise the A.I. systems.
Jared Friedman, a partner at Y Combinator, a start-up accelerator, recently said a quarter of the accelerator’s current batch of start-ups were using A.I. to write nearly all their code.”
As a programmer myself, whenever I read this I feel like there’s a giant group chat going on and I’m excluded. I use AI in my work, mostly through Zed assistant, but with the latest available models the output and reasoning is nowhere in quality where I’d let it generate majority of the code and ship it to production.
Am I missing something? Do people really generate majority of the code that then successfully operates in production at any significant business scale?
[1] - https://archive.ph/XykVf I don't, no, but that's not the most interesting thing in my workplace. The most interesting thing is that despite working for a company that produces deep learning systems, none of the devs here (that I'm aware of) are using "AI" for production code. Almost all of us have been playing around with it, but so far haven't found enough value there to use it for real. I use co-pilot for enhanced tab-completion of small blocks of code. Basically, it saves me some typing now and again. I find it guesses incorrectly a third of the time but I often still tab complete then go back and fix it. Is this "using ai to write 90% of my code"? When I have used copilot conversation mode and requested a refactoring or provided what I felt was a clear description of my task, the result shave always been a failure that required more time to resolve than if I had just done the work myself. These are issues I could see and determine where wrong. I weep for the future where these changes get pushed live because no-one knows what the vibe-coders did any more. I don't think they mean auto-complete, though I have no proof of that. If they include auto-completion, that's extremely disingenuous on their part. While I expect something like this from VCs hyping their portfolios, I'd expect NYT journalists to do more due diligence. I agree, except, co-pilot currently reports that I have used 35% of my code completion quota (and 0% of my conversation quota). I'm sure if it is tracking that telemetry, there are similar management reports that will claim it "wrote" whatever was tab completed if not claim credit for any line that included tab completion. I use it mostly as glorified auto-complete. I don’t think it’s good at replacing the creative parts of coding, but it’s very good at replacing the grunt work. I use it to write the majority of my unit tests, for example. (They of course still need tweaking, but it saves me significant amounts of time.) For some reason, any model trips up on our FE unit tests. Which are pretty generic tests for a React app. The only difference is that we use vitest instead of jest and, even with special prompts, I can't seem to make it remember that throughout the whole context window. We did some 1 month copilot test then decided not to use it. I used and still use gpt 4 for some exploratory and learning with a mixed experience here too. I ask it to populate some tests here and there or to provide debugging pointers and once to help me design a logo. Try cursor and 3.5 sonnet. It's definitely good enough. It's good if you give it specific functions to write or if you write a well structured plan in markdown.