Comment by staticman2
14 hours ago
> This behavior is exactly what you'd expect from a model distilled from Claude.
This is not at all what I would expect because it's trivial to change the training data to replace Claude with Kimi. In fact I'd argue it's almost certainly not saying that due to distillation.
I encourage you to review the links before committing to a position. The writeup on K3's anomalous trans-model identity is very comprehensive.
K3 reproduces Claude's internal model identifier when prompted, something which the real Claude models themselves do not emit. This is highly suggestive that K3 was trained on Claude metadata (API logs, tagged synthetic data), rather than Claude's chat outputs.
And it's well documented that Chinese labs are buying large amounts of raw Claude metadata https://www.chinatalk.media/p/how-to-buy-cheap-claude-tokens...
For context you just asked me to read a document that starts with:
"Caveat: fully AI-generated research."
And that you quoted or paraphrased directly.
>This is not at all what I would expect because it's trivial to change the training data to replace Claude with Kimi.
Wait what? The reason you wouldn't expect it is because if it was distilled, it would be easy to get rid of self identification? Is that any less true of a non distilled model? I suppose there's lots of ways to interpret it, but the idea that self-identifying as Claude is affirmative evidence that it's not distilled seems to get the weight of the inference exactly backwards.