Its also something that is defined differently by everyone depending on what they are selling

the third one is where it's best to stop.

sounds strange for me mix sovereignty and AI, a tool that you don’t domain, you are the user not the creator

the weights you can verify if the env is fully reproducible the local attack vectors are a problem, ideally you have a secure data center + tamper proof hardware. its not a panacea but theres a huge difference between "i'm using random llm api provider (or a proxy to have even more hops that you dont control)" and "i'm using a fully reproducible remotely attested confidential inference api" , even if the latter is not perfect for all threat models

Thanks for the clarification.

yes it does fall into private ai category. based on the claims (i havent read the source) it provides anonymization of traffic to the providers, but that addresses a very specific issue - attribution of the conversation to specific identity. it does not protect the privacy of the conversation itself, at least not from the provider. So if you upload sensitive documents or prompt it with releaving data your identity might not be tied to the account, but the data can still get out. thats where the confidential inference and confidential computing part comes in, you can get much better assurances there provided that its implemented right and you can attest to a reproducible stack that is not leaking your data somewhere.

It should probably be clearer about the limits. https://nymbot.ai/docs/anonymous/#limits

Clearer about which limits?