Benchcraft, a skeptical map of AI research

Why it surfaced

Benchcraft promises something rarer than another research roundup: a plainly stated negative verdict when the evidence does not hold up. Its method includes defining a field’s boundary, collecting preprints through a fixed date, grouping proposals by their actual approach, and sometimes running the public code while recording the commit, hardware, and failure point.

An independent survey project that reads narrow AI fields exhaustively, groups competing approaches, and judges whether they can survive outside a benchmark.