tacit/nanochat
1
0
Fork 0
mirror of https://github.com/karpathy/nanochat.git synced 2025-12-06 20:32:14 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Illustrated architecture

Browse Source
This commit is contained in:
Nda-jiya Suberu 2025-11-01 11:07:21 +00:00
parent 317e4b65df
commit e100cdf7f1
2 changed files with 66 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

24
README.md
View File

@ -1,4 +1,28 @@
# nanochat
## Extended LLM Cognitive Framework
This document outlines the new cognitive layers and features added to the nanochat LLM, along with their architectural connections.
### Key Added Features
1. **PsycheController** (in `engine.py`): Dynamically generates blending weights for the Id/Ego/Superego layers based on input context
2. **GPT Id/Ego/Superego Layers** (in `gpt.py`): Three distinct cognitive processing layers with specialized focuses (instinctual, practical, moral reasoning)
3. **HypercubeEmbeddingLayer** (in `hypercube.py`): Maps concept IDs to embeddings within a hypercube structure, enabling semantic relationships.
4. **LongTermMemory** (in `engine.py`): Stores and retrieves embeddings to preserve long-term context across generations
5. **AbacusEncoder & AbacusStateMemory** (in `abacus_encoder.py`/`abacus_state_memory.py`): Ensures logical consistency of internal reasoning through structured pattern storage
6. **MemeticLearningLayer** (in `memetic_learning.py`): Evaluates the fitness of generated embeddings for knowledge retention
7. **ConsciousIntegrationLayer** (in `conscious_integration.py`): Synthesizes outputs from all lower layers into a unified conceptual state that directly influences model output
### Architecture Flow
1. Input embeddings are processed through GPT's Id/Ego/Superego layers
2. PsycheController provides dynamic blending weights
3. LongTermMemory retrieves relevant context embeddings
4. AbacusEncoder ensures logical consistency of intermediate states
5. MemeticLearningLayer evaluates generated embedding fitness
6. ConsciousIntegrationLayer synthesizes all inputs into a unified state
7. Synthesized state modulates concept_logits for final output
### Architecture Visualization
See the vectorized diagram below for direct connections:
![LLM Cognitive Architecture](nanochat/architecture.svg)
![nanochat logo](dev/nanochat.png)

42
nanochat/architecture.svg Normal file
View File

@ -0,0 +1,42 @@
<svg width="800" height="500" xmlns="http://www.w3.org/2000/svg">
<!-- Background -->
<rect width="100%" height="100%" fill="#f8f9fa"/>
<!-- Title -->
<text x="400" y="30" font-family="Arial" font-size="20" text-anchor="middle" fill="#2d3748">LLM Cognitive Architecture Connections</text>
<!-- Nodes -->
<rect x="50" y="100" width="120" height="50" rx="5" fill="#4299e1"/>
<text x="110" y="130" font-family="Arial" font-size="14" text-anchor="middle" fill="white">Input Embeddings</text>
<rect x="220" y="80" width="120" height="90" rx="5" fill="#ed64a6"/>
<text x="280" y="120" font-family="Arial" font-size="14" text-anchor="middle" fill="white">GPT (Id/Ego/Superego)</text>
<rect x="220" y="200" width="120" height="50" rx="5" fill="#9f7aea"/>
<text x="280" y="230" font-family="Arial" font-size="14" text-anchor="middle" fill="white">PsycheController</text>
<rect x="390" y="80" width="120" height="50" rx="5" fill="#38b2ac"/>
<text x="450" y="110" font-family="Arial" font-size="14" text-anchor="middle" fill="white">LongTermMemory</text>
<rect x="390" y="150" width="120" height="50" rx="5" fill="#38b2ac"/>
<text x="450" y="180" font-family="Arial" font-size="14" text-anchor="middle" fill="white">AbacusEncoder</text>
<rect x="390" y="220" width="120" height="50" rx="5" fill="#38b2ac"/>
<text x="450" y="250" font-family="Arial" font-size="14" text-anchor="middle" fill="white">MemeticLearning</text>
<rect x="560" y="100" width="150" height="90" rx="5" fill="#f6ad55"/>
<text x="635" y="140" font-family="Arial" font-size="14" text-anchor="middle" fill="white">ConsciousIntegrationLayer</text>
<rect x="740" y="120" width="120" height="50" rx="5" fill="#48bb78"/>
<text x="800" y="150" font-family="Arial" font-size="14" text-anchor="middle" fill="white">Concept Logits (Output)</text>
<rect x="50" y="180" width="120" height="50" rx="5" fill="#4299e1"/>
<text x="110" y="210" font-family="Arial" font-size="14" text-anchor="middle" fill="white">HypercubeEmbedding</text>
<!-- Connections -->
<line x1="170" y1="125" x2="220" y2="125" stroke="#2d3748" stroke-width="2" marker-end="url(#arrowhead)"/>
<line x1="170" y1="205" x2="220" y2="125" stroke="#2d3748" stroke-width="2" marker-end="url(#arrowhead)"/>
<line x1="340" y1="125" x2="390" y2="105" stroke="#2d3748" stroke-width="2" marker-end="url(#arrowhead)"/>
<line x1="340" y1="125" x2="390" y2="175" stroke="#2d3748" stroke-width="2" marker-end="url(#arrowhead)"/>
<line x1="340" y1="125" x2="390" y2="245" stroke="#2d3748" stroke-width="2" marker-end="url(#arrowhead)"/>
<line x1="340" y1="225" x2="390" y2="125" stroke="#2d3748" stroke-width="2" marker-end="url(#arrowhead)"/>
<line x1="510" y1="105" x2="560" y2="125" stroke="#2d3748" stroke-width="2" marker-end="url(#arrowhead)"/>
<line x1="510" y1="175" x2="560" y2="125" stroke="#2d3748" stroke-width="2" marker-end="url(#arrowhead)"/>
<line x1="510" y1="245" x2="560" y2="125" stroke="#2d3748" stroke-width="2" marker-end="url(#arrowhead)"/>
<line x1="710" y1="145" x2="740" y2="145" stroke="#2d3748" stroke-width="2" marker-end="url(#arrowhead)"/>
<!-- Arrowhead Marker -->
<defs>
<marker id="arrowhead" markerWidth="10" markerHeight="7" refX="9" refY="3.5" orient="auto">
<polygon points="0 0, 10 3.5, 0 7" fill="#2d3748"/>
</marker>
</defs>
</svg>
Side by Side

After

Width:  |  Height:  |  Size: 3.3 KiB