How Pythia Works

How Pythia is designed to allocate compute across the network. Today Pythia runs as a standalone service; this describes the plan.

The Decision Engine

Pythia is designed to evaluate network conditions and decide how to allocate compute resources across the mesh.

1

Sense

Pythia continuously monitors network state: task queue depth, energy costs, and node availability across all tiers.

2

Predict

Using Active Inference, Pythia generates predictions about future states and calculates expected value for each allocation strategy.

3

Act

The action that minimizes free energy (surprise) is selected. Compute is allocated to the best available use.

Active Inference

Built on Karl Friston's Free Energy Principle, Active Inference is a framework from computational neuroscience that describes how intelligent agents minimize surprise in their environment.

  • ●Generative Model: Pythia maintains a model of how the network behaves and uses it to predict outcomes.
  • ●Belief Updating: As new data arrives, beliefs are updated using Bayesian inference.
  • ●Policy Selection: Actions are chosen to minimize expected free energy (surprise + uncertainty).
// Simplified Active Inference Loop
while network.is_running() {
  // 1. Observe current state
  state = sense_network();

  // 2. Update beliefs
  beliefs = update_beliefs(
    prior_beliefs,
    state
  );

  // 3. Evaluate policies
  policies = [
    allocate_to_tasks(beliefs),
    allocate_to_other(beliefs),
    hybrid_allocation(beliefs)
  ];

  // 4. Select action minimizing
  //    expected free energy
  action = argmin(
    policies.map(p =>
      expected_free_energy(p, beliefs)
    )
  );

  // 5. Execute and learn
  execute(action);
}

Compute Allocation

Planned: Pythia would evaluate how available compute is best used, such as running user tasks or other workloads.

Task Execution Mode

When the task queue is full and prices are favorable:

  • • AI inference workloads
  • • Data processing jobs
  • • Model training (federated)
  • • Custom compute tasks
Rewards
Not yet defined

Idle Mode

When task demand is low (not yet defined):

  • • Other uses not yet defined
  • • Efficient power management
Rewards
Not yet defined
1
Local Training
Data stays on your device
2
Gradient Encryption
Only encrypted updates shared
3
Secure Aggregation
Pythia combines learnings
4
Model Distribution
Improved model sent to all

Federated Learning

Pythia is designed to learn from the network without seeing raw data. Each device trains locally; only encrypted gradients are shared.

  • ✓Raw data never leaves your device
  • ✓Differential privacy protects individual contributions
  • ✓The shared model is designed to improve over time

Explore the Technology Stack

Dive deeper into the cryptographic protocols, network topology, and system architecture that powers Pythia AI.