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...

>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.

You've backtracked significantly here.

"Zero evidence" as you claimed earlier isn't accurate. You've moved the goalposts from "evidence" to "raw internal logs I can independently audit," which is a different and very high standard. Sure Anthropic didn't publish logs, IP addresses, timestamps, or account IDs of the accounts involved. But that's true of any cybersecurity breach/abuse disclosure ever made. Companies are furtive to reveal how they detect fraud, because doing so exposes the signals used to detect bad actors, and makes future abuse easier. Not revealing the "evidence" you're asking for is industry standard practice. You're complaining that Anthropic is following industry standard practice, and conveniently defining the "evidence" you need as something Anthropic is never going to publish.

> I don’t consider a tweet by Denise Wu, who works at Anthropic, to be reproducible evidence.

Is the issue here that she works at Anthropic? Because Denise Wu doesn't work there.

> I don’t consider “Caveat: fully AI-generated research” to be someone taking time to analyze anything in great detail.

The experiments were run by Ryan Greenblatt, who is a real AI safety researcher (at Redwood Research).

The identity experiments and Greenblatt analysis are trivially reproducible. The methodology, code, and metrics are all there in the Github repository. You can ask your preferred AI to independently replicate these results, and it will give you a result within an hour.

You’ve also reduced the evidence to “two models producing vaguely similar front-end styles,” which is not what either analysis shows.

From the analysis, Kimi K3 identifies itself as Claude 15% of the time. How do you explain that? Qwen and GPT identify themselves as Claude 0% of the time.

If a long document is too much analysis for you, someone else made a simple chart which measures the KL divergence between Kimi K3 and other major models. They found K3 is unusually similar to Fable 5 & Opus models. That is, Kimi K3 has an very similar style and phrasing to that of Anthropic models. That behavior is expected from a model distilled from Claude.

https://typebulb.com/u/lab/you-re-relatively-right/full

What's in the base model is "bits and pieces of intelligence".

You have to bring those bits and pieces together, put them into the right shapes and fill in the gaps to get a model that actually performs. This is what post-training is all about. It's not at all a trivial thing.

Reasoning, tool use, agentic behavior - all of those are post-training performance gains. Getting a good well trained base model is putting your foot in the door of frontier performance - post-training is how you actually get inside.

See: GPT-4.5 vs o1. One went for "build a bigger better more capable base model", the other went for "take the old base and post-train it for advanced capabilities". The results: a wider base with basic post-training loses to a narrower base with advanced post-training. Or, hell: GPT-3 vs GPT-3.5. One was largely a research lab curio, and the other kicked off the AI revolution as we know it.

The gains compound. Getting a better base model with the same type of post-training helps, see: the jump from Opus to Mythos/Fable. But post-training techniques account for a lot of the performance juice.

And yes, reasoning trace post-training distillation is "genuine distillation". As is logit distillation in pre-training. "Distillation" isn't a single training recipe that you have to follow to a tee - it's a large group of training methods. I've seen plenty of wacky things like inverse distillation bootstrap and post-training self-distillation that use distillation in strange ways at different stages of the training run to get results.

Do I really have to explain the shape of AI training pipelines to you?

Train a big, wide base model with a lot of potential. Mid-train or post-train that on Claude Opus 4.5 reasoning/agentic traces (i.e. Claude Code data from Chinese API resellers) to make your model approximate a high baseline of chatbot behavior, reasoning, agentic work and tool use.

Then run your own expensive SFT, RLHF and RLVR on top of that yoinked baseline to dial it in further.

Actually doing RLHF and RLVR is extremely expensive. Distillation gives you a lot of dense, high quality post-training signal for cheap. This can get your model into the basin of "the right way to tackle this kind of problem" without a frontier lab compute budget. It's a big shortcut that gets you closer to the target - you can take it from there and build on top of it with your own work.

Also, it's unclear whether "summarizing thinking tokens" actually ruins distillation, or just makes it work worse. I'd bet on the latter, really. Because it's an approximation game, and summarized reasoning is still a better approximation of true reasoning than most of what you get online and in pre-training datasets.

You're conflating pre-training data volume with post-training data volume.

Nobody is suggesting Moonshot used 15 trillion tokens of Claude data to pre-train a base model from scratch. That would be impossible and nonsensical.

This is entirely about distillation, which happens during post-training (alignment and SFT). Here, datasets are measured in millions or billions of tokens, not trillions. 50 billion Claude tokens is far, far than enough to copy Claude's reasoning logic, writing style, and tool-use ability to the pre-trained base model.

> However, this is not distillation

I don't understand how you're so caught up on the term "distillation". Distillation is using a larger model's outputs to train a (weaker) student model. Which is exactly what's happening. It's a standardized term that has been in use for a decade.