Скачать или смотреть The Unexpected Effectiveness of One-Shot Decompilation with Claude

The blog post explores the successful implementation of "one-shot" decompilation using the large language model Claude in an unattended, continuous loop, dramatically accelerating progress on the Snowboard Kids 2 project. This automated system bypasses the need for constant human supervision, relying on a four-part workflow: a scorer that prioritizes the easiest functions, Claude (primarily Opus 4.5) which performs the function matching, a toolbox of defensive Unix-like programs, and a driver that manages the entire lifecycle, failure handling, and logging,,,. This method proved highly efficient, demonstrating that LLMs could potentially match about 79% of the project's functions. This approach shifts the primary constraint on traditional, multi-year decompilation efforts away from the time of human experts and toward compute power and access to frontier models. Nevertheless, the role of human experts remains vital for handling the most challenging functions, developing the necessary tools, and cleaning up the sometimes "ugly" code generated by Claude, which often contains awkward constructs like pointer arithmetic or excessive `goto` statements.

https://blog.chrislewis.au/the-unexpe...