Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

MiniMax M2.5 is trained by Claude Opus 4.6?

news.ycombinator.com

Post reply on HN
Print view View original

11–15 of 15 posts

Re: MiniMax M2.5 is trained by Claude Opus 4.6?

#11

Post by orbital-decay »

Balinares wrote: post #10
orbital-decay wrote: post #7

Earlier quoted context omitted.

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.

What makes these cases unambiguous?

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...

Re: MiniMax M2.5 is trained by Claude Opus 4.6?

#14

Post by KomoD »

Well yeah, that is what Anthropic is claiming so I'm not surprised.

https://www.nytimes.com/2026/02/23/technology/anthropic-chin...

https://www.anthropic.com/news/detecting-and-preventing-dist...

> We have identified industrial-scale campaigns by three AI laboratories—DeepSeek, Moonshot, and MiniMax—to illicitly extract Claude’s capabilities to improve their own models. These labs generated over 16 million exchanges with Claude through approximately 24,000 fraudulent accounts, in violation of our terms of service and regional access restrictions.

Re: MiniMax M2.5 is trained by Claude Opus 4.6?

#15

Post by Balinares »

orbital-decay wrote: post #11
Balinares wrote: post #10

Earlier quoted context omitted.

What makes these cases unambiguous?

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…

Interesting! Thanks for taking the time to write down the explanation, I appreciate it.
Post reply on HN