Skip to content

Peleke/aegir

Repository files navigation

Aegir

The jötunn who personified the sea and hosted feasts for the Æsir

A self-paced curriculum for learning the mathematical foundations of concept spaces and variational reasoning — toward building physics-inspired AI memory and reasoning systems.

The Vision

What if reasoning paths through concept space followed the principle of least action? What if concepts emerged from experience rather than being predefined? What if justification was intrinsic to the geometry rather than external metadata?

This curriculum builds the mathematical foundation to explore these ideas.

Quick Start

# Clone and enter
cd ~/Documents/Projects/aegir

# Install dependencies (requires uv)
uv sync

# Launch JupyterLab
jupyter lab

# Start with notebooks/00-setup/00-environment-test.ipynb

Structure

aegir/
├── wiki/           # Primers on core concepts
├── sources/        # Curated reading list
├── syllabus/       # Week-by-week curriculum
├── notebooks/      # Hands-on learning (Jupyter)
├── lib/            # Reusable Python code
├── data/           # Sample data
└── experiments/    # Research logs

Curriculum Overview

Week Topic Milestone
0 Onboarding First embedding visualization
1 Foundations Gradient descent from scratch
2 Embeddings Mini Word2Vec working
3 Information Theory Compute MI between clusters
4 Clustering Name emergent concept clusters
5 Dynamical Systems Visualize attractor landscape
6 Variational Calculus Derive shortest path
7 Differential Geometry Compute geodesics on sphere
8 Geometric Deep Learning Build simple GNN
9 Integration Working concept space prototype
10 Experiments Research results

Design Principles

ADHD-Optimized

  • Micro-modules (30-45 min each)
  • Visual progress tracking
  • Quick wins early
  • Interleaved theory/practice

Hands-On First

  • Every concept → immediate code
  • Never more than 10 min reading before doing
  • Visualize everything

Build Toward Something Real

  • Not academic exercises
  • Culminates in working prototype
  • Designed for multi-source RAG integration

Prerequisites

  • Python fluency
  • Basic familiarity with linear algebra/calculus (we'll refresh)
  • Curiosity about the nature of concepts and reasoning

The Big Idea

We're building toward a system where:

  1. Concepts emerge from density in embedding space
  2. Salience creates a potential field that shapes attention
  3. Reasoning is movement through this space
  4. Paths minimize an action functional (least effort + maximum relevance)
  5. Justification is intrinsic — the path exists because it's optimal

Think of it as Lagrangian mechanics for thought.

Resources

See sources/ for the full reading list. Key resources:

  • 3Blue1Brown - Linear algebra, calculus intuition
  • Strogatz - Nonlinear Dynamics and Chaos
  • Bronstein et al. - Geometric Deep Learning
  • Raschka - Build an LLM from Scratch

License

MIT

About

Recursive curriculum generation system with four-layer pedagogy

Resources

Stars

Watchers

Forks

Packages

 
 
 

Contributors