Resources
Glossary
The terms used across PRINET.
- Activation
- The intermediate values a model computes at each layer. In a sharded network, they are what passes between nodes.
- Chance baseline
- How much of a prompt an attacker would recover by guessing: 0.2%. A run is private only if a node does no better.
- Entry / exit node
- The first and last nodes on a request's path. They sit closest to the tokens and see the most.
- Open weights
- A model whose parameters are public. Anyone, including an attacker, can run it.
- Recoverability
- The share of prompt tokens an attacker can reconstruct from one node's view.
- Secret share
- One of two pieces a value is split into. Each alone is noise; only both together recover the value.
- Sharded inference
- Serving a model by splitting its layers into blocks that run on different machines.
- Swarm
- A group of contributed GPUs that together serve every layer of one model.
- 2PC
- Two-party computation: two parties compute on secret shares without either seeing the real data.
- Worker
- A node that lends GPU compute to the network and gets paid per token served.