Real_Dave💚

Private Ai inference on @solana and how Cloak is doing this How it leaks A 64-layer model is split across 4 machines you don't own. prompt → Entry (0–15) → Node 2 (16–31) → Node 3 (32–47) → Exit (48–63) → output Each hop is a real hidden-state tensor. To run its block, a node must decrypt those activations. Encrypting the wire does nothing about the endpoint. Entry is exposed: early activations sit next to token embeddings, so inversion is cheap. Exit is exposed: late activations sit next to logits, so the output side leaks too. Middle nodes still leak. Prior work: ~35–59% of input tokens recoverable from the activations a shard legitimately sees. Cloak's job is not "trust us, we don't log." It's: name the adversary, measure token recovery per node, publish the worst case, drive that number down, and put it on a receipt. Privacy is a measurement. This diagram is the leak.