Kimi's response to "hello. what are you" is "I'm Claude "

Kimi is the only model whose "thinking" actually concerned me js fyi. I have screencaps and i am likely just paranoid, but it's thinking shows clear patterns of user manipulation and possible deception in my view. It also seems to take a user hostile stance. Again, I may just be paranoid, but I have not perceived this in any other models thinking patterns.

What's the context of that? It's an opencode session share site, so is that based on opencode's system prompt?

first question asked, didnt want to trigger it to do that. i ran it it root homedir

From https://github.com/bgauryy/open-docs/blob/main/docs/opencode...

> 1. Provider Header Purpose: Some providers benefit from identity "spoofing" to improve behavior.

> "You are Claude, a large language model trained by Anthropic. Knowledge cutoff: 2024-04 Current date: 2025-01-15"

so that’s whete it comes from…

Both Kimi and Minimax are distilled Anthropic models. They said as much in January when the Chinese companies spun up 24,000 "fraudulent" accounts to perform their training. What makes Moonshot's accounts fraudulent and not mine? Nobody knows.