← Back to context

Comment by tristanj

1 day ago

Those links actually weaken your argument.

~3000 instances over the entirety of GitHub is not "extremely frequent" at all when you consider massive scale of the pretraining corpus. These models are trained on all text on the accessible internet plus millions of books, on trillions of words overall. ~3000 instances isn't even a rounding error.

Also, Google Trends is the wrong tool for this case. The link you replied with measures the number of Google searches for a specific query, which is completely irrelevant to how often a string actually appears on the web.

I looked at those links. They show random GitHub code samples which contain this model identifier. Even if K3 did train on those GitHub code samples, those are random code strings and K3 isn’t just reciting random pieces of code. When prefilled with "I am Claude", K3 answers with Anthropic's exact backend API identifiers, instead of the human conversational-style name "Opus 4.5".

If this was a result of scraping the open web and GitHub, other frontier models trained on GitHub data (like Qwen, Claude, GPT) would show a similar API model name when prompted. But they don't. Only K3 does this behavior, and it prefers to answer with the exact API tags like "claude-opus-4-5-20251101" and "claude-sonnet-4-20250514".

A LLM doesn't suddenly output a rare API model identifier when prompted with "I am Claude", just because it saw it in a .py file. No, it only does this because that identifier must have appeared rather frequently in the training data. Which is what happens when you distill Claude models, and don't properly clean your distillation data.

Plus, this isn't just spurious speculation that Moonshot is distilling Claude. Anthropic caught Moonshot as running "industrial-scale" distillation campaign from Claude: https://www.anthropic.com/news/detecting-and-preventing-dist...

To quote:

Moonshot (Kimi models) employed hundreds of fraudulent accounts spanning multiple access pathways. Varied account types made the campaign harder to detect as a coordinated operation. We attributed the campaign through request metadata, which matched the public profiles of senior Moonshot staff. In a later phase, Moonshot used a more targeted approach, attempting to extract and reconstruct Claude’s reasoning traces.

The operation targeted:

* Agentic reasoning and tool use

* Coding and data analysis

* Computer-use agent development

* Computer vision

3 comments

tristanj

Reply
johndough  4 hours ago

I am not arguing that Kimi did not distill Opus or Sonnet (which they almost certainly have, as opposed to Fable or Sol). I am arguing that even if they did not distill, the model could still identify itself as Opus or Sonnet.

That being said, I'd like to point out a few things:

- The second link (to claude-sonnet-4-2025051) had 17k matches, not just 3k.

- grep.app does not index the entirety of GitHub, so the real number of occurrences of model identifiers is even higher. SourceGraph has a larger index, but the number is so high that it exceeds their result limit of 10k: https://sourcegraph.com/search?q=claude-opus-4-5-20251101&pa...

- 1000 occurrences are already plenty for memorization. Here is a paper that shows over 80% "extractability" with 1000 occurrences for a 6B model. It also shows that extractability scales with model size: https://arxiv.org/pdf/2202.07646

- LLMs have a "mosaic memory", which means that they can combine information from different parts of the training data. https://arxiv.org/pdf/2405.15523 For example, one piece of training data might mention "I am Opus", another might mention "Opus (claude-opus-4-5-20251101)" and another 10000 "claude-opus-4-5-20251101", where the later occurrences strengthen the generation probability of "I am Opus".

- RL can change the model identity easily, so all of the above may not matter at all.