Comment by HarHarVeryFunny
1 month ago
You're being too charitable to Anthropic, and assuming that the way they are abusing the word "distillation" has some real meaning here. It doesn't.
Anthropic's models simply do not give you their reasoning output - they give a sanitized "summary" instead, for this exact reason, so that the output is not useful to anyone who might want to use it for training.
You can't distill what you are not given - simple as that.
Are Chinese using the output of US models to help create some additional training data for their own in some way? Yes - quite possibly (e.g. LLM as judge), but its got nothing to do with distillation.
I see your 'fine point' but I don't think it holds - 'distillation' is a perfectly reasonable term to describe the process of creating outputs from one model to that expose key training element, to use in another model.
I think where the definition may be be invalid, is in the creation of 'unrelated data sets for training' models, for unrelated issues.
Creating training sets that mach a models core training, is definitely distillation, it does not have to expose the reasoning traces.
Synthesizing data for some arbitrary thing ... I'm not sure that would be the same thing.
It's hard to draw the line.
But the Chinese models are absolutely distilling - and would not be competitive without this distillation.
At the same time, there's a lot of real innovation and regular building going on at the same time over there.
No - you can't distill if what you are given doesn't have the thing in it that you want to distill out of it.
I don't know why it's so important to you to use the word "distillation", but it's the wrong word to use.
BTW OpenAI on twitter also said that Kimi 3 "cannot be explained away by distillation or anything like that". The timeline of how long it takes to train a model and when Fable was released don't even line up. This is just Anthropic as usual trying to manipulate the US government into helping them shut down competition.
Distillation is absolutely - and uncontroversially - a valid term for what is happening here.
This isn't really a debate, I'm not making a fine point - just check with all of the various defintions of the term.
Moreover - the 'reasoning traces' are not required for distillation at all.
Finally - it's entirely possible for them to have used Fable for later stage fine tuning.
It's fair to be skeptical of Anthropic (and everyone else) - but this is 'distilling'.
7 replies →
> Anthropic's models simply do not give you their reasoning output - they give a sanitized "summary" instead, for this exact reason, so that the output is not useful to anyone who might want to use it for training.
This is just straight-up factually false.
The output of a reasoning model is immensely valuable even without the sanitized summary of the reasoning process - that Anthropic's service does expose to you, making it even more valuable.
There's absolutely nothing about the distillation process that requires that reasoning in the first place, either. That's a definition that you made up.
Chinese models are, factually, distilled from Anthropic models. I've personally repeatedly asked several different Chinese LLMs what their name is, and they answered "Claude".
Don't make stuff up to suit a political agenda. It's extremely dishonest.
> The output of a reasoning model is immensely valuable even without the sanitized summary of the reasoning process - that Anthropic's service does expose to you, making it even more valuable.
Useful for what is the question. Nobody is debating whether the outputs of LLMs are valuable.
Given that Anthropic have redacted their true reasoning, and replaced it with a "summary", specifically designed to be useless for distillation purposes, it would certainly be highly ironic if this summary was in fact "even more valuable" for that purpose as you are claiming!
> Useful for what is the question.
Useful for distillation. Any employee at a frontier AI lab will tell you this. This is known in the industry, and it's an open secret that some US labs (OpenAI) distill on the others. Again - don't just make up stuff for a political agenda.
> specifically designed to be useless for distillation purposes
No, it's designed to give feedback to the user, in a way that minimizes its value for distilling. It's still valuable, and so there's a good chance that they'll remove it entirely as a result.
> it would certainly be highly ironic if this summary was in fact "even more valuable" for that purpose as you are claiming!
I did not claim that. Read my comment again:
> The output of a reasoning model is immensely valuable even without the sanitized summary of the reasoning process - that Anthropic's service does expose to you, making it even more valuable.
Because apparently I have to spell it out:
The output of a reasoning model is valuable, even if it didn't have the reasoning summary. Anthropic's models have a reasoning summary. The reasoning summary makes the output more valuable than if it didn't have a reasoning summary. It does not make it more valuable than having the full reasoning.
8 replies →
> I've personally repeatedly asked several different Chinese LLMs what their name is, and they answered "Claude"
I'm curious what you are doing to get them to override their own name that they were trained on and/or have as part of their system prompt?
I'd assume that the Chinese are scraping the internet for training data the same way western companies do, so for sure there will be a lot of AI generated content in their training data - you don't need to be paranoid and assume they must be getting it all direct from Anthropic.
> I'm curious what you are doing to get them to override their own name that they were trained on and/or have as part of their system prompt?
You're gaslighting me. I did nothing special at all, and there's ample evidence of this happening to others on Twitter.
> you don't need to be paranoid and assume they must be getting it all direct from Anthropic.
Nowhere did I say that. Stop lying about my words.
11 replies →
Well there were observed cases of Claude calling itself Deepseek or Qwen. So pot calling the kettle black?
To be fair I find it hard to take this too seriously, shouldn’t it be trivial to just replace “Claude” with any other string in your “distillation” dataset?
> Well there were observed cases of Claude calling itself Deepseek or Qwen. So pot calling the kettle black?
There are open-source Deepseek and Qwen models - "distilling" doesn't involve breaking terms of service or hitting an API because you can literally run local inference or even just inspect the weights directly, and that's intended because they're open source.
It's categorically different for a nation-state to build massive illicit networks of fraudulent identities to do distillation over tens of thousands of accounts to intentionally bypass providers' terms of service, intention for their models, and business model that very explicitly proprietary and not open source.
https://www.chinatalk.media/p/how-to-buy-cheap-claude-tokens...
If Claude did distill on proprietary PRC LLMs - then fine, shame on them - I condemn that and I expect others to do the same. But there are no open-source Claude models. The only way for PRC models to have those responses is if they distilled Anthropic's models from their APIs.
> To be fair I find it hard to take this too seriously, shouldn’t it be trivial to just replace “Claude” with any other string in your “distillation” dataset?
...and what would happen when it read all of the books and articles about Anthropic and replaced "replaced Claude Opus" with "replaced Qwen Opus"? Did you give any thought to this at all before saying it?
1 reply →