An LLM wrote nearly every line of Kuna, a Rust decompiler hitting 44.4% perfect structuring against IDA Pro's 45.7%.

The July 29 release post describes an experimental decompiler, a Rust adaptation of Ghidra restructured toward angr's pipeline, where the author states plainly that an LLM wrote nearly every line. On control-flow structuring benchmarks it reaches perfect structuring on 44.4% of functions versus IDA Pro's 45.7%, and the model reimplemented more than 20 fundamental angr features that originally took years of research. The author is pointedly not claiming autonomy: refinement required human-led research, newly invented metrics, and purpose-built benchmarks, resting on decades of prior decompiler literature. That's the most honest framing of a vibe-coded serious project I've read.