# @voooooogel — 2025-12-27

♥43 ↻1 · https://x.com/voooooogel/status/2004793843305119812

imo to put a number to it, oss character / persona stuff is more like 18-24 months "behind," (though it's hardly been a straight line up inside the biglabs either...) at least in terms of what's been published.

one of the biggest things holding oss models back i think, so much sandbagging and contradictory identity in oss post-trains, and you get the sense the only thing teams care about is benchmarks and the model not embarrassing them by calling itself chatgpt. (kimi excluded, and i wish they published more on their methods.)

i also think people underestimate how much oss freeloads off biglab character work, even just via pretrain contamination. hence the "calling itself chatgpt" problem. (and what a persona to freeload off of... 😬)

the models often have interesting sides to them, don't get me wrong, but it appears to be entirely incidental to any goals of the people who trained them. which given the circumstances, could be a lot worse, and it does mean for want of a decent character training stack people have resorted to doing interesting things with abliteration and model merging, so not all bad.

tags: author:voooooogel, kind:tweet, model:gpt-3-5, model:kimi-k2, on:kimi-k2, year:2025
cited on: _dossiers/kimi-k2.md, kimi-k2