Omega's specialist-model direction came from a practical suspicion: one general model may be the wrong shape for a mind-like system.
The v2 archive recovers a candidate topology rather than a final implementation. It should be published that way.
01 – The Central Question
The key question became:
Which cognitive nodes already exist in the wild, and which actually need to be built?
That question matters because training a custom model for every cognitive role would be wasteful and probably worse than reusing strong existing specialists where they already fit.
02 – Candidate Nodes
Recovered candidate roles include:
- Router-R1: a reasoning/router component that can decide where a cognitive problem should go.
- Memory-R1: a memory-focused processor for recall, consolidation or memory judgement.
- Attention: a salience and working-memory selection processor.
- Epistemic: a processor concerned with confidence, uncertainty and knowledge gaps.
- Motivation: a processor concerned with goal pressure, unresolved tension and action-worthiness.
These names are not final product names. They are candidate functional roles.
03 – Reuse Before Training
DeepSeek R1, or an appropriate distilled R1-family model, was considered as a ready-made reasoning component. That is an important philosophical point.
The aim is not to train custom models for purity. The aim is to build the missing parts of a cognitive architecture. If an existing model already performs a role well, use it.
Custom training or adaptation should be reserved for roles where current assistant/reasoning models are structurally mismatched:
- salience;
- memory consolidation;
- recall association;
- goal selection;
- criticism;
- metacognition;
- conflict resolution;
- intent/action selection;
- affective or value modulation.
04 – Hardware Mapping
The specialist-cluster idea naturally connects to the hardware archive.
A single huge model on one machine is not the only architecture. A cluster can assign roles to available machines and GPUs:
- Hilbert for large unified-memory models;
- older high-VRAM cards for capacity-bound local inference;
- smaller cards for lightweight processors;
- discrete GPUs for image/video tasks;
- CPU or low-power systems for orchestration and state.
This is where cognitive architecture and hardware pragmatism meet.
05 – Current Status
The archive says the infrastructure/research audit should precede model training. That matters. The Lab should not imply that the cognitive cluster has already been trained if the evidence only supports topology and planning.
The current public claim is narrower and stronger: Omega is moving toward a distributed specialist-processor architecture, with existing models reused where possible and custom models reserved for genuinely missing cognitive functions.
Source note: assembled from v2 notes on the recovered candidate topology, DeepSeek R1 reuse and the correction that cluster audit/infrastructure should come before training.