FilterHN

MiniMax M2.5 is trained by Claude Opus 4.6?

10 points

17 hours ago

| 6 comments

I was chatting with MiniMax M2.5 in OpenRouter and suddenly he mysteriously repeated on "I'm Claude, an AI assistant created by Anthropic - not a "language" ", heh wut?

▲

evan-dong

13 hours ago

[-]

You really can’t trust what an LLM says about its own identity. If it has seen tons of chats saying “I’m Claude / GPT‑4 / DeepSeek”, it will just echo that pattern in similar contexts. That’s more about dataset contamination and pattern matching than any real evidence it is that model.

▲

throwaw12

12 hours ago

[-]

Isn't it great news for us?

You get an open model which is a 95% of Opus 4.6 quality and 80% cheaper in most inference providers and also can run on your own hardware

Also they did the hard parts of:

* crawling the content

* running the fine tuning (or training)

Better than 1 or 2 companies taking control of the whole AI economy

▲

muzani

16 hours ago

[-]

They all are trained by each other. Claude says it's DeepSeek if you ask it in Mandarin.

▲

SilverElfin

16 hours ago

[-]

Most people seem to think that phenomenon is not the same thing. People have shown by experimenting with different prompts that even in Mandarin, Claude correctly says it’s Claude when it is doing something for you. But if you ask it about its identity, it sometimes says DeepSeek. The current theory is it just has run into Chinese content that has chat logs that often have a DeepSeek model answering that it is DeepSeek. But the inconsistency in different prompts suggests this is something different from distillation.

▲

pogue

13 hours ago

[-]

Is theft of theft theft?

▲

orbital-decay

14 hours ago

[-]

If it was, that line is not an indicator. Distillation is done on useful prompts, not on "Who are you?" - "I'm this model of that company".

Name training is always shallow, Claude itself would claim it's GPT-3, GPT-4, or Reddit (heh) when confused. It's just dataset contamination, because the web is full of slop. Never trust self-reported names.

▲

SilverElfin

16 hours ago

[-]

This has been a common issue with the Chinese open weight models. It appears most or all have been trained via distillation on OpenAI and Anthropic models.

▲

orbital-decay

13 hours ago

[-]

They most likely weren't, despite very dubious claims of Amodei and Altman and a certain twitter influencer running a pretty naive writing benchmark ("slop test") that is wrong in a very obvious manner. The only unambiguous cases of distillation were Gemini 2.0 experimentals being trained on Claude outputs, and GLM-4.7 being trained on Gemini 3.0 Pro. The rest are pretty different from each other.

▲

Balinares

9 hours ago

[-]

What makes these cases unambiguous?

▲

orbital-decay

6 hours ago

[-]

GLM-4.7 (specifically this version) repeats the guardrail prompt injections from 3.0 Pro, word-by-word, and never follows them, which is consistent with training on a reward-hacked CoT. Gemini 3.0 only discusses snippets from this injection in its native CoT (hidden by default, trivial to uncover), but GLM-4.7 was able to reconstruct it in full during training. The only possible reason for this is direct training on a large amount of examples of Gemini's CoT. Its structure and a lot of replies were identical in GLM too.

Gemini 2.0 Exp 1206 was reported to be indirectly trained on Claude's outputs with humans in between [1], which was pretty consistent with its outputs at the time. No other Gemini versions except two experimental ones were similar to Claude.

[1] https://techcrunch.com/2024/12/24/google-is-using-anthropics...