Tell HN: Kimi AI builds persistent profiles from your conversations

3 points

4 months ago

I've been using Kimi sporadically over the last month since they offered a free tier, and I purchased a subscription today when they had a $1 first-month deal. During a discussion, something tipped me off, so I checked the model's thinking process.

    The user is a senior dev based on their profile (Rust, Unix, etc.), so I should speak to that level of experience rather than entry-level advice. They work in Rust, systems programming, security-focused tools - these are actually harder to automate areas than CRUD web apps.
    Their profile shows they:
    - Run Arch Linux (power user)
    - Work with Rust (systems programming - harder for AI to do safely)
    - Build security-focused tools
    - Do financial research (JPY/CNH hedging, geopolitics, energy markets)
    - Are direct/blunt communication style
    - Prefer open source, drop-in replacements, structured outputs

The thinking tokens don't reveal whether previous conversations were loaded in that moment or if a profile had been built over time. I examined the conversation titles (particularly the one about JPY/CNH hedging, which came up deep within a discussion). Based on this (https://www.reddit.com/r/LocalLLaMA/comments/1qoml1n/leaked_kimi_k25s_full_system_prompt_tools/), it seems like they synthesize a profile from your conversations.

Just something to be aware of...

1 comment