Kuna is an experimental decompiler developed with nearly all code written by an LLM, rivaling IDA Pro in control flow structuring with 44.4% perfect functions versus IDA's 45.7%.
The decompiler improves through autonomous refinement by studying examples where it underperforms against IDA Pro, Ghidra, and angr, reimplementing over 20 features from angr.
Kuna is a Rust port of Ghidra adapted to angr's pipeline, and it relies on decades of decompiler research from scientists and engineers.
Scientific insight and human-led research are essential for Kuna's automatic refinement, including discovering new metrics and meaningful benchmarks.
GPT-5 failed to count the number of 'b's in 'blueberry', answering 3 instead of 2, sparking criticism on Bluesky.
The failure is partly due to tokenization, where LLMs process text as tokens rather than individual letters, making letter counting difficult.
Author conducted a systematic test of multiple LLMs, finding that most models correctly answered the 'strawberry' question (likely memorized), but GPT-5 Chat frequently failed on 'blueberry'.
Other models like Claude Sonnet 4 and Opus 4.1 performed perfectly, while Gemini 2.5 Flash and open-source models showed inconsistent errors.