Codex
Each sanctuary dramatises a real idea. Here they are without the metaphor, with a link to where the idea comes from. Read these in any order, before or after the pilgrimage.
- Axiom of Openness
The Glass House
Nous argues for glass-house development: open weights and inspectable methods, so that anyone can audit how a model reaches an output rather than trusting an opaque service.
Transparency is not decoration. Inspectability lets people test, contest and improve a system.
Nous Research - Axiom of Coordination
The Psyche Constellation
Psyche coordinates model training across heterogeneous, unreliable hardware: an onchain coordinator sets the schedule, clients train, and witnesses attest that work was really done.
Distributed hardware can cooperate when coordination, onboarding and verification are designed in.
Psyche - Axiom of Signal
The DisTrO Choir
DisTrO and DeMo compress what training nodes must exchange by orders of magnitude, making distributed training viable over ordinary internet links.
Useful coordination can depend on sending the important structure rather than every raw detail.
DisTrO - Axiom of Transformation
The Forge of Questions
DataForge composes synthetic data as a graph of transformations, instruction generation, answering and judging, so each stage can be inspected and replaced.
Data synthesis can be compositional and reviewed, shaped for usefulness rather than produced as a mass.
Nous open source - Axiom of Learning
The Atropos Observatory
Atropos is an RL environment framework: it runs trajectories, scores them, and feeds verified outcomes back so behaviour can actually change.
Improvement needs environments, trajectories and verification — not punishment, but selection.
Atropos