As someone who doesn't use AI for writing code, why can't you just ask Claude to write up an explanation of each change for code review? Then at least you can look at whether the explanation seems sane.
Because the explanations will often not be sane; when they are sane, they will focus on irrelevant details and be maddeningly padded out unless you put inordinate effort into trying to control the AI's writing style.
Ask pretty much any FOSS developer who has received AI-generated (both code and explanations) PRs on GitHub (and when you complain about these, the author will almost always use the same AI to generate responses) about their experiences. It's a huge time sink if you don't cut them off. There are plenty of projects out there now that have explicit policy documentation against such submissions and even boilerplate messages for rejecting them.
It will fairly confidently state changes are "correct" for whatever reason it makes up. This becomes more of an issue with things that might be edge cases or vague requirements, in which case it's better to have AI write tests instead of the code.
This can be dangerous, because Claude doesn't truly understand why it did something. Whatever it writes a post-hoc justification which may or may not be accurate to the "intent". This is because these are still autoregressive models --- they have only the context to go on, not prior intent.
Indeed. Watching it (well, Anthropic, really) cheat at Baba Is You and then try to give a rationalization for how it came up with the solution (qv. https://news.ycombinator.com/item?id=44473615) is quite instructive.
Claude also doesn't know, because Claude dreamt up changes that didn't work, then "fixed" them, "fixed" them again and in the process left swathes of code that isn't reached.
I've been experimenting with Claude, and feel like it works quite well if I micromanage it. I will ask it: "Ok, but why this way and not the simpler way? And it will go "You are absolutely right" and implement the changes exactly how I want them. At least I think it does. Repeatedly, I've looked at a PR I created (and review myself, as I'm not using it "on production"), and found some pretty useless stuff mixed into otherwise solid PRs. These things are so easily missed.
That said, the models, or to be more precise, the tools surrounding it and the craft of interacting with it, are still improving at a pace where I now believe we will get to a point where "hand-crafted" code is the exception in a matter of years.
AI is not a human. If it understands things it doesn't understand things like you or I. This means it can misunderstand things in ways we can't understand.
AI is not sentient, so it does not “understand” anything. I don’t expect the autocomplete of my messenger app to understand its output, so why should I expect Claude to understand its output?
