Million Minds

Creating a novel AI framework from the ground up based on neuroscience principles, exploring biologically-inspired intelligence systems

Concept

Million Minds is an ambitious project aimed at developing a novel AI framework inspired by the fundamental principles of neuroscience. Unlike today's AI systems that rely heavily on deep learning and large datasets, Million Minds seeks to emulate the brain's architecture and functionality to create more efficient, adaptable, and intelligent systems.

The current prototype starts with text: an eye-like sensor, cortical columns, lateral cooperation, and an early hierarchy.

Neuroscience Foundations

  • Brain is a sensorimotor, space processor
  • Brain operates in analog; no bits, no clocks, no messages
  • Brain has a repeating cortical structure
  • Each structure is running the same algorithm
  • Neurons don’t understand what they are doing, they leverage:
    • Sparse encoding
    • Homeostasis
    • Co-firing and association (Hebbian learning)
    • Predictive state
    • Oscillation
    • Layering

Benefits

  • True intelligence:
    • Builds direct, indirect and abstract world models
    • Learns patterns, concepts, relationships
  • Can learn at any time, learns continuously
  • Learning does not require retaining everything
  • Needs much less data to train
  • Uses far less memory, compute, energy to run
  • Modality agnostic, can work with text, image, audio, video, sensor data and more

Bottom Line

By grounding AI development in the principles that govern biological intelligence, Million Minds is designed to overcome the limitations of current AI technologies, such as their lack of generalization and adaptability. This approach has the potential to achieve breakthroughs leading the way to AGI and beyond.

Latest Notes

  • June 26, 2026
    I have been building a small step-through demo so I can watch the system read text: the sensor moves, Layer 1 settles, Layer 2 reacts, and I can…
  • June 12, 2026
    Ambiguity is not always failure. That was the main lesson from the latest round of work.The recent focus was Layer 1 and early Layer 2 behavior over…
  • June 5, 2026
    Layer 1 is starting to work like an object-recognition layer over text.The system scans text through small sensor patches, then cortical-column-like…

Roadmap

Built So Far

  • ✅ Text Vision
    The system can scan text with an eye-like sensor instead of treating documents as a plain token stream.
  • ✅ First-Layer Learning
    The first layer can learn recurring local text patterns, recognize familiar ones, and become less surprised when it sees them again.
  • ✅ Column Cooperation
    Multiple columns can observe overlapping views and exchange context, laying the groundwork for a many-model system rather than one centralized model.
  • ✅ Early Hierarchy
    A second layer is running and can learn from the stable output of the first layer.
  • ✅ Inspection Tools
    There are tests and an interactive demo that show the system training, replaying familiar text, making predictions, and detecting changed input.

Current Focus

  • Stacking More Layers
    The next step is to make the hierarchy deeper, so higher layers can learn larger patterns from lower layers.
  • Stronger Recognition
    The system needs to become more reliable at recognizing familiar patterns while keeping genuinely ambiguous possibilities alive.
  • Better Cooperation
    Columns should learn stronger associations with each other so context from one view can help another view.
  • Practical Speed
    The architecture needs to run faster so it can handle more columns, deeper layers, and larger text collections on normal hardware.

Later Directions

  • Top-Down Understanding
    Higher layers should eventually help lower layers interpret ambiguous input, similar to how context shapes perception.
  • Learned Attention
    The system should eventually decide where to look next instead of following a fixed scanning pattern.
  • Native Language Generation
    The long-term goal is for the system to generate language from its own learned predictions, not rely on a transformer to speak for it.
  • More Modalities
    The same architecture should eventually extend beyond text to images, video, 2D layouts, virtual environments, and hardware sensors.

Last Updated: Jun 28, 2026