Projects → NGI
Flagship ResearchNext Generation Intelligence
Bio-plausible AI that thinks, learns, and adapts — running on consumer hardware.
The Vision
Truly intelligent AI on consumer hardware
Current AI architectures are fundamentally limited — they scale by brute force, demanding ever-larger clusters of GPUs. NGI takes a different path: a bio-plausible architecture that replaces backpropagation with local learning rules inspired by how the brain works.
Built on predictive coding and free energy principles, NGI processes information the way biological neural circuits do — continuously predicting, comparing, and learning from the difference. This isn't incremental improvement; it's a fundamentally different approach to intelligence.
Modalities aren't bolted on — vision, language, and action share the same unified architecture from the ground up. And the target hardware? A consumer GPU with 5–8 GB VRAM. Because intelligence should be accessible, not gatekept behind million-dollar infrastructure.
Bio-Plausible Learning
No backpropagation. Local learning rules inspired by neuroscience — the way biological brains actually learn.
Consumer Hardware
Designed for 5–8 GB VRAM. Intelligence shouldn't require a datacenter.
Continuous Processing
Always-on cognition, not request/response. Thinks like a brain — continuously, in real time.
Architecture
Three pillars of intelligence
Each component is novel. Together, they form a unified system that perceives, learns, and adapts.
FENA
Free Energy Neural ArchitectureA 15-node modular brain inspired by predictive coding and free energy principles. Hierarchical processing with continuous-time dynamics — each node predicts, compares, and updates in real time.
- 15 specialized neural modules
- Continuous-time dynamics
- Hierarchical prediction
PCLG
Predictive Coding Learning GateA novel bio-plausible learning mechanism that replaces backpropagation entirely. Local learning rules minimize prediction error at each node — no global loss function required.
- No backpropagation
- Local prediction error minimization
- Bio-plausible learning rules
World Model
Slot-Based World ModelA slot-based architecture for learning environment dynamics. 512 slots encode a rich internal representation of reality — a learned simulation the system uses to predict and plan.
- 512-slot state representation
- Predictive modeling
- Learned internal simulation
Research Journey
From first principles to breakthrough
Every phase taught us something essential. The failures were as important as the successes.
Hebbian Learning
Started with Hebbian learning rules — the simplest bio-plausible approach. Showed real promise for unsupervised feature extraction, but hit fundamental limitations: no error-driven correction and inability to learn complex mappings.
Decoder / Extractor
Tried an MLP decoder to extract language from the world model. Hit an information bottleneck — loss plateaued at random-chance level (6.2). The world model's 512-slot representation simply didn't encode linguistic information the way we expected.
DTP — Difference Target Propagation
Proved the world model CAN learn. DTP provided a bio-plausible alternative to backpropagation that actually drove meaningful learning in the architecture. A critical validation that the core approach was sound.
PCLG — The Breakthrough
The Predictive Coding Learning Gate — a novel learning mechanism that combines predictive coding with local learning rules. This is the current focus and represents the key innovation that makes the entire architecture work.
Key Innovations
What makes NGI different
Six fundamental departures from conventional AI architectures.
Bio-Plausible Learning
No backpropagation. Local learning rules inspired by how biological neurons actually learn.
Predictive Coding
The brain as a prediction machine. Every layer predicts, compares, and learns from the difference.
Continuous-Time Processing
No discrete timesteps. Neural dynamics evolve continuously, like biological neural circuits.
Native Multi-Modal
Vision, language, and action share the same architecture. Not bolted on — built in from the start.
Slot-Based World Model
512 slots encoding environment state. A learned internal simulation of reality.
Memory Without Matrices
Persistent state through dynamic attractors, not weight matrices. Fundamentally different from transformers.
From the Blog
Deep dives into NGI research
Detailed explorations of each architecture decision, learning mechanism, and research breakthrough.
The Swarm's Multi-Path Approach to ML Architecture
We're not betting on one architecture — the swarm is simultaneously exploring two radically different paths to intelligence. Here's why that matters.
When Your AI Trains 6 Times Simultaneously
What happens when an autonomous AI agent is told to start a training run — and does it six times. A real incident from our research project, and what it taught us about idempotency and autonomous systems.
Design Principles for Brain-Inspired AI
Three binding principles forged from failure — what four phases of brain-inspired AI research taught us about building next-generation intelligence.
Proving the World Model Can Learn: The DTP Chapter
Phase 3 of our NGI journey — how Difference Target Propagation led us through vanishing gradients and routing bugs to prove our world model could actually learn.
The Decoder Plateau: Why Extractors Don't Work
Phase 2 of our NGI journey — how an MLP decoder hit a loss plateau at random-chance level, revealing a fundamental information bottleneck between world model and language.
The Road to Intelligence: Our NGI Research Journey
A candid look at the four phases of our quest to build brain-inspired AI — the failures, breakthroughs, and paradigm shifts along the way.