Skip to content

Overview

Separate cognition from execution

The GNUS.ai platform has two complementary concerns:

  1. Genius Cognitive System (GCS): A controller/planner selects a Semantic Core and task-appropriate Expert Language Models (ELMs), assembles relevant GAML memory under privacy and retention rules, mediates tools, and checks or arbitrates results.
  2. GNUS / SuperGenius: Provides the execution, storage, peer-to-peer, proof, ledger, and settlement capabilities for eligible distributed jobs. A GCS request need not use the public network; local and private execution are supported architectural modes.

This is not conventional agent orchestration plus a sharded large model. ELMs are specialized models or roles selected per task; GAML is governed, structured memory rather than unbounded conversation replay. Answer quality checks and execution-integrity checks are separate concerns. The complete cognitive design is documented at gcs.gnus.ai.

flowchart TD
    Client[Application or API] --> GCS[GCS controller / planner]
    GCS --> Core[Semantic Core]
    GCS --> Experts[Selected Expert Language Models]
    GCS --> Memory[GAML: governed memory]
    GCS --> Checks[Grounding, arbitration, and answer checks]
    GCS --> Execution{Execution choice}
    Execution --> Local[Local or private runtime]
    Execution --> Swarm[GNUS / SuperGenius jobs]
    Swarm --> P2P[Peer-to-peer compute and storage]
    Swarm --> Verify[Execution verification]
    Swarm --> Settle[Ledger and settlement]

Infrastructure building blocks

The earlier distributed retrieval and processing design uses:

  • RocksDB and IPFS for storage, with CRDT-based synchronization where appropriate.
  • Vulkan, ggml, and MNN for supported local processing and inference paths.
  • libp2p pub/sub for distributed communication.
  • Retrieval and training workflows for workloads that need distributed RAG or model updates.

These components describe possible execution paths, not a guarantee that every GCS feature or distributed workflow is already deployed or that results are deterministic across all GPU vendors. For current release status see Platform Status; for subsystem ownership see Master Architecture.

Reading the rest of this section

The Query Workflow, Data Storage, Pub/Sub Communication, and Retraining Mechanism pages cover individual infrastructure designs. They should not be read as a complete description of GCS reasoning or memory.