The research-tooling bottleneck just moved

The bottleneck for research-grade software just moved. For decades, the limiting factor in serious scientific tools was human implementation time: every algorithm hand-designed, hand-coded, hand-tuned. That assumption is breaking, and the breakage lives in feedback loops, not raw model capability.

The receipt is in noelo.org's Kuna release. Kuna is a decompiler, software that translates compiled machine code back into something a human can read, released as an experiment in instrumented scientific feedback. A language model wrote nearly every line. The model was not given a specification and told to generate. It was given a closed loop: examine cases where Kuna underperforms established tools, study how those tools solved the hard ones, refine against the gap. Run the loop, and the model reimplemented, in Kuna, more than 20 fundamental features from angr, the decompiler the author helped build over years of scientific work.

That is a different claim than "AI wrote a decompiler." A single model run is not a research tool. A loop that ingests its own failures, studies a more capable reference, and reinvents its missing features is closer to a slow, narrow version of the scientific process itself. Ghidra and angr remain the substrate; angr remains the long-term home for frontier algorithms; Kuna is the harder-to-hand-write experiment that runs on top.

For research-grade decompilers, the bottleneck may shift toward how well the feedback loop is instrumented, rather than how many engineers a lab can hire.