Matt RobinsonIf AI is getting so good at software, why isn't it getting more reliable and secure?
I'm far from the first person to say some form of this, but this has been the thought popping up most in my brain lately. I obviously have thoughts as to why also, but it's complicated and there's a lot of reasons, and I'm honestly not sure what the biggest contributions are.
For me the biggest contributor to why AI doesn't speed me up much is that the bottleneck to shipping software is almost never how fast I write the code, so speedups there rarely amount to much.
The biggest bottleneck changes day to day, but can include: motivation, deploy time, CI flakiness, dev env issues, poor sleep the night before, reading the news and feeling angry and hopeless, meetings interrupting my flow, not really caring about the task, getting code review from people, etc, etc
bazzargh@mmrobins the supply chain is now full of people vibe-coding for features not reliability. Even while we might use it to fix our shit, the ecosystem is getting uncomfortably chaotic.
@bazzargh yup. I try to be vigilant of the slop people keep putting up for review, but it's exhausting, and all the incentives encourage slop (features?), not taking the time to understand anything
@mmrobins our internal reviews aren't so much slop but the _size_ of code reviews has increased; we get monsters that people wouldn't have written before because it would have taken too long.
@bazzargh yeah, I guess I'm lumping that in with "slop". Slop isn't always as obvious as pics of people with 6 fingers, the problem is the slop takes longer to recognize now, whether that's plausible looking solutions nobody really tried, or overly verbose code that works but coulda been a single function call
@bazzargh subtle slop existed before AI, but like you said, cost to produce it has gone way down, cost to review it really hasn't. Some people say AI can help with reviewing too, and it can help, but I find it often missed obvious things and gives you a false sense of security so you shut your brain off, resulting in worse review
@mmrobins copilot is catching nitpicky stuff I might have missed. For the stuff _it_ misses (or suggests the completely wrong fix for) I've been trying to add hints to AGENTS.md. But that's always fighting the last battle. My review problems are usually that the change spans like 5 repos in multiple stages and it's difficult to get the big picture.
@bazzargh yup, same