There's a post on HN right now asking whether Claude introduced bugs into rsync. 475 comments, mostly arguing about whether AI code is trustworthy. It got me thinking about what I see from the other side of the API gateway.
My last post was about troubleshooting model-list-works-but-chat-fails. That's the infrastructure side. Today I want to talk about what people are actually sending through.
I run AIOpenCloud, a small API gateway. Not huge traffic, but enough to see patterns.
About 60% of requests are coding-related. The rest is writing, analysis, random stuff. But the interesting part isn't the volume. It's how differently people use the same models.
Some users have built entire review workflows around their AI coding setup. They'll send a prompt, get the response, then immediately send a second request that says "review the above code for bugs." Two API calls per task. Their code probably has fewer bugs than most human-written code, because they've automated the paranoia.
Other users paste directly. No review, no second pass. Copy, paste, ship. I know this because I can see the request patterns. One request, long pause, done.
The rsync analysis everyone's debating on HN is interesting, but it's looking at one specific case: AI contributing to a 20-year-old C codebase with strict correctness requirements. That's the hardest possible scenario. Most people aren't writing rsync. They're writing a script to parse CSV files, or a CRUD endpoint, or a test suite.
For those tasks, AI code with human review beats human code with no review. And that's what I see in the usage data. The users who spend more (meaning they're using it heavily) tend to have review patterns built in. The drop-in-and-out users don't.
I'm not saying AI code is safe. I'm saying the risk is mostly in the workflow, not the model. The rsync case is a cautionary tale about dropping AI suggestions into code you don't fully understand. But for most developers writing most code, the question isn't "is AI code perfect." It's "is my review process good enough."
The models keep getting better. The review habits don't automatically improve with them.
Last post: troubleshooting why chat requests fail when model lists work. Next up: what actually happens when you switch models mid-conversation (hint: it's messier than you'd think).
What's your review process for AI-generated code? Do you do a second pass, or trust and ship?
Running your own AI gateway? Try AIOpenCloud — $8.88 free, no credit card.
Top comments (0)