←back to thread

491 points todsacerdoti | 2 comments | | HN request time: 0.411s | source
1. daeken ◴[] No.44382870[source]
I've been trying out Claude Code (the tool I've found most effective in terms of agentic code gen/manipulation) for an emulator project of mine for the last few days. Part of it is a compiler from an architecture definition to disassembler/interpreter/recompiler. I hit a fairly minor compiler bug and decided to ask Claude to debug and fix it. Some things I noted:

1. My C# code compiled just fine and ran even, but it was convinced that I was missing a closing brace on a lambda near where the exception was occurring. The diff was ... Putting the existing brace on a new line. Confidently stated that was the problem and declared it fixed.

2. It did figure out that an unexpected type was being seen, and implemented a pathway that allowed for it to get to the next error, but didn't look into why that type had gotten there; that was the actual bug, not the unhandled type. So it "fixed" it, but just kicked the can down the road.

3. When figuring out the issue, it just looked at the stack trace. That was it. It was running the compiler itself; it could've just embedded some debug code (like I did) and work out what the actual issue was, but it didn't even try. The exception was just a NotSupportedException with no extra details to work off of, so adding just a crumb of context would let you solve the issue.

Now, is this the simplest emulator you could throw AI at? No, not at all. But neither is qemu. I'm thoroughly unconvinced that current tools could provide real value on codebases like these. I'm bullish on them for the future, and I use GenAI constantly, but this ain't a viable use case today.

replies(1): >>44390085 #
2. lowbloodsugar ◴[] No.44390085[source]
This is the part that makes me sure my job is safe. I use AI to write code, but it really sucks at debugging.