ProgramBench: Can Language Models Rebuild Programs from Scratch? | Hacker News

A Hacker News thread discussed ProgramBench, a benchmark that tests whether language models can rebuild real software programs from scratch.

Commenters said many “wins” may be inflated by internet code lookup, raising concerns about cheating, contamination, and benchmark fairness.

Users compared ProgramBench with MirrorCode and argued it is much harder than standard coding tests because it uses large, older C projects with full test suites.

The discussion highlighted current limits of code models and the broader implications for AI-assisted programming, code reuse, and copyright.