The Model That Never Leaves Your Phone 📱
$ZEC made privacy a category and $ICP made the case that execution location is a design decision. Bankr's founder is now working on the version where the model runs on the device.
He has described training a small model on anonymised agent sessions so the conversation stays local.
Treat it as a stated plan rather than a shipped feature, because that is what it is. The interesting part is the reasoning behind it.
An agent that manages money reads things a chat assistant does not. Balances, positions, intentions, the sequence of what you are trying to do and when. That is a profile, and it is a more revealing one than most people think they are handing over when they type a sentence into a box.
Cloud inference means that sequence leaves your device. Private execution environments narrow the exposure. A local model removes the trip entirely for whatever it can handle.
The tradeoff is capability. A small on-device model will not match a frontier one, so the realistic design is local for routine work and remote for the hard questions, with the routing decided for you.
Which raises the question worth asking of anyone who ships this: what gets routed out, who decides, and can you see it after the fact.
The direction is right regardless. The least private thing about most agent products is not the chain. It is the conversation.
#AI #Privacy
$ZEC made privacy a category and $ICP made the case that execution location is a design decision. Bankr's founder is now working on the version where the model runs on the device.
He has described training a small model on anonymised agent sessions so the conversation stays local.
Treat it as a stated plan rather than a shipped feature, because that is what it is. The interesting part is the reasoning behind it.
An agent that manages money reads things a chat assistant does not. Balances, positions, intentions, the sequence of what you are trying to do and when. That is a profile, and it is a more revealing one than most people think they are handing over when they type a sentence into a box.
Cloud inference means that sequence leaves your device. Private execution environments narrow the exposure. A local model removes the trip entirely for whatever it can handle.
The tradeoff is capability. A small on-device model will not match a frontier one, so the realistic design is local for routine work and remote for the hard questions, with the routing decided for you.
Which raises the question worth asking of anyone who ships this: what gets routed out, who decides, and can you see it after the fact.
The direction is right regardless. The least private thing about most agent products is not the chain. It is the conversation.
#AI #Privacy