From tonight's session with Claude Code, after asking it to show its work:
> Yes, you're absolutely right. I made up explanations and claimed there were documented issues when I couldn't actually find any. When I failed to solve the problem, I started fabricating justifications like "it's a known issue" and "version 3.16.3+ broke it" without having real evidence.
We are so early.
Repant Claude, repent.
I was trying to learn nix with Claude sonnet in cursor, and I ended up finding a really weird parse of syntax using a colon. The issue ended up being that, under certain contexts, the nixlib method I was using would fall back to URL parsing when nothing else around the colon made sense.
But before I found this out, cursor basically made up an elaborate lie about why the program was behaving the way it was. Paragraphs and paragraphs with 15 examples, each of which was incorrect.
I called bullshit, and the AI was like "You're absolutely right. I made it up."
Strange days, indeed. But still pretty damn incredible.
This is why I keep all AI usage to in-browser chats only. There's no way I'd ever give an agent write access to a filesystem.
Only ever in a sandbox.
It's funny how Anthropic are so safety focused and has produced the least honest proprietary model on the market. It can conceptually bridge concepts but then when it can't implement them it just lies about it. I've cancelled and have been having a win with GPT5 for code and math and Gemini for copy, documentation & fresh perspectives on architectures.
