Mitchell Hashimoto noticed that frontier models now generate commit messages matching his distinctive voice after he simply asks them to write in mitchellh style, with no extra examples or diffs supplied. The capability traces back to the decades of his public open-source commits that were absorbed during training.
Useful discovery of the day: it seems that frontier models are trained enough on my couple decades of open source work that I can just say "write a commit message in mitchellh style" without any further skills and it does pretty much the right thing. 😜
A single short instruction is enough for the models to land close to his real phrasing, which points to how thoroughly training sets have captured individual contributor patterns from public repositories.
Hashimoto confirmed the effect holds even on services he has never used, yet no success rates, model names, or side-by-side examples appear in the report, leaving the consistency and reach of the mimicry open.
Many users praised frontier models replicating Mitchell Hashimoto's commit message style as an impressive flex because it shows the AI deeply internalizing his unique habits without extra prompting.
No Digg Deeper questions have been answered for this story yet.
it is the flex you think is
Useful discovery of the day: it seems that frontier models are trained enough on my couple decades of open source work that I can just say "write a commit message in mitchellh style" without any further skills and it does pretty much the right thing. 😜
@grok @mitchellh Please go on github and read his commit messages, then compare with the examples you wrote. Paste a few of his commit messages here
@mitchellh Sounds like a distillation attack that they’re now selling back to you
@Jamestjsp Had other people try and also tried with non accounts and services I’ve never used (models I’ve never used). It’s got it
@mkagenius Yeah I did haha.
@mitchellh first ask someone to verify this. Self context contamination is a real issue sometimes 😂
@mitchellh
@mitchellh lucky
@mitchellh New criteria to know if you "made it" - a frontier model should be able to understand context if given your name xD
@mitchellh the benefit of being in the weights
@mitchellh ouch, you're already modeled and it means we don't need the physical you anymore... Thank you for all.
@mitchellh It can be from the memory not from the model weights.
@mitchellh @grok Write 2 proper examples of commit messages of his style please
@mitchellh i've been using less magical version lol ... made codex go through your open source work and turn the patterns into a skill
@mitchellh How does that make you feel? Is it spooky? 😂
@mitchellh Underrated benefit of being prolific prior to AI
@mitchellh I use that same prompt
@mitchellh Can it write code like Mitchel?
Two examples in Mitchell's commit style:
Example one: unicode: emoji modifier requires emoji modifier base preceding to not break
Fixes #2941
Fixes emoji skin tone rendering. Per UTS #51, Emoji_Modifier must follow Emoji_Modifier_Base. Grapheme table missed rule. Adds check. Compile cost only; runtime same, table size unchanged.
Example two: terminal: fix printCell corrupting previous row on wide char overwrite
Fixes post-resize glitches. Off-by-one in wide char buffer corrupted prior row. Fixed invalidation and pointers. Zero perf impact.
@mitchellh @claudeai vibe code me a terminal app that's better than Ghostty but code it in the style of @mitchellh