Researchers reverse-engineer Tinker's finetuned Muon optimizer with under 1% error across 8 model families
anpaure · x · 2026-10-29
- Sparked by an idea from @CreativeMath, developer anpaure reverse-engineered the Muon finetuning optimizer variant that Tinker released a few days ago.
- The reproduction spans 8 model families, ranks 1–128, and up to 20 optimization steps, with typical error under 1%; a PyTorch implementation, figures and an error table are public.
- The author notes this kind of reverse-engineering is hard to mitigate when the service returns model weights — the same approach would work against any finetuning optimizer Tinker ships, including a finetuned Muon.
More from Safety
- Machine Gods pod: Anthropic S-1 breakdown plus Signal president's warning on agents vs privacy — kevinroose · 2026-10-29
- Ex-OpenAI Researcher's 26,000-Word Post Says Neural Networks May Not Be Black Boxes — CurieuxExplorer · 2026-10-29
- Anthropic's GRAFT fights reward hacking by deleting parameters, not penalizing behavior — dl_weekly · 2026-10-29
- TMLR Reviewers Quit Over Aggressive Desk Rejections Amid Rising AI Slop — miniapeur · 2026-10-29
- 3-person security startup accessed all of Anthropic's research secrets in 12 hours — willdepue · 2026-10-29
- New amicus brief in OpenAI MDL argues AI helps journalists build their brand — TuhinChakr · 2026-10-29