Loading…
14
Hot take: A guy at a 2019 meetup in Austin said AI would never write code that ships, and I just shipped a fix it wrote
Back in 2019 I sat through a meetup in Austin where this older dev stood up and said machine learning would never touch real production code, that it was all autocomplete with extra steps. I nodded along because at the time every code helper I tried spat out junk that broke the build. Fast forward to last Tuesday, I had a null check failing at 2am on a service that handles about 40k requests a day, and the AI tool I use now spotted the edge case, wrote the patch, and it passed CI on the first run. I sat there staring at the green check thinking about that guy. The tools now read the whole repo, not just the line you are on, and that was the thing nobody saw coming back then. I still read every line it gives me, no way am I merging blind. But the gap between 2019 autocomplete and what I ran at 2am is not a small step, it is a different sport. Anybody else have an old prediction about AI that aged badly, or one that actually held up?
1 comments
Log in to join the discussion
Log In1 Comment
henderson.vera1mo ago
That line about "a different sport" is doing a lot of heavy lifting for what might just be a better autocomplete. One patch passing CI at 2am is one data point, and im sure we have all pushed changes that passed CI and still broke something a week later. The tool read your repo, sure, but it read a snapshot of the mess you already made, it did not understand why that null check existed or what upstream caller is going to trip over the new behavior. Green check just means the tests you have did not catch it, and half the nasty bugs live in the edges nobody wrote a test for. So maybe the old guy was wrong about the when, but he was not crazy about the why.
2