Check yourself: Think without phone-home

Questions and answers, readable without JavaScript.

1. What is local inference, as this note defines it?

   • Weights on hardware you control, generating tokens on that machine or LAN, unless you wire the prompt path out yourself
   • A SaaS API with a local-looking theme
   • A public swarm that shares activations
   • A model card downloaded without the weights
   Answer 1

   Weights on hardware you control, generating tokens on that machine or LAN, unless you wire the prompt path out yourself

   Local inference means you load weights onto hardware you control and generate tokens there. Nothing in the prompt path needs to leave unless you wire it out.

   Read the note → · Local and Peer-to-Peer AI: Data Residency Without the Hype

2. What does local residency remove?

   • Malware and shoulder-surfing
   • The need to verify weights
   • Provider exposure for that session — not the need for a secure operating system
   • Every copy of the prompt, including disk
   Answer 2

   Provider exposure for that session — not the need for a secure operating system

   There is no third-party training clause for that local session. A compromised laptop with a local model is still a compromised laptop.

   Read the note → · Local and Peer-to-Peer AI: Data Residency Without the Hype

3. What is a model card, in this note?

   • A signature over every tensor
   • Proof the file was not tampered with
   • A licence to run the model in the cloud
   • Structured publisher claims — intended use, training notes, limits — that you can check, not a security guarantee
   Answer 3

   Structured publisher claims — intended use, training notes, limits — that you can check, not a security guarantee

   Cards document intended use, training notes, evaluation, and known limits. Treat them as publisher claims you can check against checksums and your own evals, not as a guarantee.

   Read the note → · Local and Peer-to-Peer AI: Data Residency Without the Hype