The Long Tail of LLM-Assisted Decompilation
4 hours ago
- #decompilation
- #coding agents
- #Nintendo 64
- Coding agents were used to decompile Nintendo 64 games, with one-shot decompilation proving highly effective.
- Initial progress on Snowboard Kids 2 decompilation jumped from 25% to 58% matched code quickly.
- Progress slowed after reaching ~75% matched code, requiring workflow adjustments.
- Function prioritization based on difficulty worked early on but stalled when only hard functions remained.
- Function similarity via text embeddings and tools like Coddog improved decompilation by providing references.
- Specialized tooling, such as gfxdis.f3dex2 for graphics microcode, enhanced Claude's effectiveness.
- Decomp-permuter was initially used but later removed due to introducing noise and instability.
- Code cleanup and documentation improved reference quality for similarity-based scheduling.
- Workflow scaling was achieved through worktrees, automation, hooks, and task orchestration with Nigel.
- GLM (glaude) was introduced as a cost-effective alternative to Opus for less demanding tasks.
- The decompilation project faced challenges with remaining functions, requiring more attempts and resources.
- The Snowboard Kids 2 decompilation project is ongoing, with community involvement encouraged.