On Device Works Until It Does Not 📱
The phone makers answered this first, and the answer was to run what fits locally and send the rest to a server, which means the hard questions are exactly the ones that leave.
That is the honest limit of on device AI. A small model handles the easy half, and anything that needs scale, memory or a frontier model goes to hardware you do not own.
$SOL apps face the same split as agents start holding keys and paying for things, because a program has to be handed a whole input before it can act on it, and the interesting inputs are the ones nobody wants to send.
So the guarantee people think they have at the device boundary disappears at the exact moment the task gets valuable.
$ARX applies to the half that leaves. An input is split into fragments across a cluster of nodes where no single node holds a readable copy, and the correct answer still comes back.
That answer settles on Solana as an ordinary public transaction, so the app behaves exactly as it did before.
Mainnet Alpha has run since February 2, and Blackthorn, which brings this to AI models themselves, has not shipped yet.
Keeping the easy work on the phone was never the hard part, and the version worth waiting for is the one where the work that leaves is still unreadable when it arrives.
#AI #Solana
The phone makers answered this first, and the answer was to run what fits locally and send the rest to a server, which means the hard questions are exactly the ones that leave.
That is the honest limit of on device AI. A small model handles the easy half, and anything that needs scale, memory or a frontier model goes to hardware you do not own.
$SOL apps face the same split as agents start holding keys and paying for things, because a program has to be handed a whole input before it can act on it, and the interesting inputs are the ones nobody wants to send.
So the guarantee people think they have at the device boundary disappears at the exact moment the task gets valuable.
$ARX applies to the half that leaves. An input is split into fragments across a cluster of nodes where no single node holds a readable copy, and the correct answer still comes back.
That answer settles on Solana as an ordinary public transaction, so the app behaves exactly as it did before.
Mainnet Alpha has run since February 2, and Blackthorn, which brings this to AI models themselves, has not shipped yet.
Keeping the easy work on the phone was never the hard part, and the version worth waiting for is the one where the work that leaves is still unreadable when it arrives.
#AI #Solana