A model that solves a problem 25% of the time is unreliable. Run it 16 times and keep the best result, and it's reliable. The math works. The bottleneck is compute.
Phage removes the bottleneck by running all 16 attempts on 16 different GPUs at once. Wall-clock time stays the same. Success rate goes up.
Volunteers. Install phage-node, contribute your idle GPU. The daemon runs sandboxed inference and submits results. Stop it whenever you want.
The coordinator doesn't care if your card is a 3060 or an A100. It dispatches based on VRAM and load.
Most distributed compute is passive. Submit work, wait, get results.
Kell is active. It breaks goals into subtasks, dispatches them, evaluates results, decides what to try next. Nodes drop and it reassigns. New nodes appear and it uses them. It monitors its own capacity and adjusts.
BOINC distributes predefined work units. Phage generates its own tasks from high-level goals. Every node runs LLM inference specifically. Results are verified by running tests, not by trusting the output.
Early. Coordinator, node daemon, gRPC protocol, verification, and model loading all work. Tensor parallelism, speculative decoding, autonomous task generation, and dashboard are planned.
Install the daemon if you want to contribute.