Mastodawn

there's a "clean room" re-implementation of claude-code in rust. it claims it was a two-step process:
1. analyze claude-code to produce a text spec.
2. implement the spec without reference to the original code.

this is a lie. the rust code has copied things in the original that aren't in the spec. in particular, the ascii art for the "buddies" are not in the spec, but the rust strings are identical to the original.

unclear if it's the human or the coding agent that lied.

Show thread

felix (grayscale) 🐺4d ago

the "clean room" implementation has noticeably better "architecture" than the original. this suggests a stupid "refactoring" loop:

- starting from a codebase P
- prompt agent to generate an extensive testsuite T that achieves 100% coverage
- prompt agent to generate a text spec S
- prompt agent to implement S without using P, such that it passes T
- the result is your new codebase P'
- throw out the old T and S and repeat until the codebase is perfect

Show thread

mirth

@gray17 I didn't read what they did but the Claude models are certainly trained on Claude Code output. The whole thing is a bit comical.