Comment by jdthedisciple
20 hours ago
> Kimi will train their models on your interactions
I find these kinds of concerns increasingly silly: most of the input to these models will be ... previous output from the very same models, alongside the occasional half-assed human command to fix something and "make zero mistakes". Who cares if they train on that? Let them, if it makes their future models better!
99% of users are not working on any special IP to worry about that.
It means there's a non-trivial chance a future version of the model will know private information about you.
Maybe you're super careful with this stuff, but with agents and harnesses being given access to user data and accounts, I don't think it's feasible to actually monitor what information is uploaded and whether they involve private information.
I personally keep local models around because of this.
I run my agents in a VM. They don't even have access to their own LLM API keys.