If a top-tier model like Claude can misrepresent its own identity because of contamination, what does that say about relying on any LLM's self-reporting for critical tasks?
I find it interesting. My own testing over six months with several Chinese models (under the Hermes update) showed something different. Every time i restarted the session and asked “What model are you using?” the Chinese models consistently gave the correct answer.
And I'm STILL waiting for any actual evidence from Anthropic to prove any such "distillation" by Chinese labs,
Not that I even care. Distillation in that form is the same as downloading copyrighted files from the Internet - it levels the playing field and is unconstrained by contract. The Chinese have every reason to do it and they should if they can. But they probably don't for the simple reason that it's less likely to be effective than other methods.
And given that Anthropic LIES all the time and on top of that explicitly wants Chinese models to be banned, because they know that eventually Chinese models will be as good as any US frontier lab, I'd say the evidence lies on the other side.
I feel this arguemnt absured that due to data contamination an LLM can misidentify itself. Its like watching many movies I would identify myself as a movie character, thats not believable.
If a top-tier model like Claude can misrepresent its own identity because of contamination, what does that say about relying on any LLM's self-reporting for critical tasks?
I find it interesting. My own testing over six months with several Chinese models (under the Hermes update) showed something different. Every time i restarted the session and asked “What model are you using?” the Chinese models consistently gave the correct answer.
And I'm STILL waiting for any actual evidence from Anthropic to prove any such "distillation" by Chinese labs,
Not that I even care. Distillation in that form is the same as downloading copyrighted files from the Internet - it levels the playing field and is unconstrained by contract. The Chinese have every reason to do it and they should if they can. But they probably don't for the simple reason that it's less likely to be effective than other methods.
And given that Anthropic LIES all the time and on top of that explicitly wants Chinese models to be banned, because they know that eventually Chinese models will be as good as any US frontier lab, I'd say the evidence lies on the other side.
I feel this arguemnt absured that due to data contamination an LLM can misidentify itself. Its like watching many movies I would identify myself as a movie character, thats not believable.
It's a rookie mistake to use any web crawled data without spending enough time to ensure it's not contaminated. What a joke.