6 hours ago
- Anthropic's clean room experiment for a C compiler in Rust had questionable methodology (no ISA docs, Rust choice, missing research papers).
- Author conducted his own clean room experiment: write a Z80 emulator, then ZX Spectrum emulator, then CP/M emulator using Claude Code.
- Process: write specification markdown, gather documentation via internet, then start fresh session with strict clean room rules (no internet, no similar source code).
- For Z80 emulator: zero steering; agent worked 20-30 min, produced 1200 lines of readable C code passing ZEXDOC/ZEXALL tests.
- For ZX Spectrum: used extensive steering for TAP loading; agent implemented quickly with SDL example, ran Jetpac with sound.
- CP/M environment implemented easily; agent inspected COM files and implemented syscalls.
- Lessons: provide agents with design hints and documentation; think like a human; LLMs assemble knowledge rather than just memorizing and copying.
- Author believes this contradicts the idea that LLMs only copy training data; the code is new and not a copyright violation.
- Next step: compare with experiment without any documentation provided to the agent.