From a medical professional with 30 years in patient care
who spent years simulating biological systems in C++
From a virtual heart that beat with 12 leads of electrical data,
responding to drugs and modeled conduction abnormalities... to biomimetic robotic fingers and hands...
He was always building toward an electrical mind. He just didn't know it yet.

Coming Soon

Evolved Neural Circuit

Intelligence that grows.

A new kind of AI — evolved from random noise, trained without backpropagation, capable of learning continuously after deployment. Not an incremental improvement to existing models. A fundamentally different path toward artificial general intelligence.

Patent Pending

Why It Runs Faster on CPU → Read the Technical Whitepaper → View the License → Live Demo — Coming Soon

What if intelligence wasn't designed — but evolved?

Every biological mind on Earth — every human, every octopus, every crow that bends wire into a hook — arrived at general intelligence the same way: 3.8 billion years of evolution. Random mutation. Selection pressure. No architect. No blueprint. Just relentless competition and the ones that worked got to keep going.

We can run that same process in silicon — compressed from eons into right now. Start with random bytes. Apply selection pressure: predict human language, or be replaced. Distinct computational species emerge from noise. They organize into layered circuits with reward, punishment, and permanent memory.

Every AI system in production today was designed by humans, trained by gradient descent, and frozen at deployment. It can never learn another thing. The Evolved Neural Circuit takes the other path — the one that actually produced intelligence the first time around. The result is a system with properties that no transformer possesses: online learning, neuroplasticity, biological reward mechanisms, and the ability to grow with every interaction. It learns the way brains learn. And it never stops.

Where it stands today

This is early-stage technology with demonstrated results. A working circuit — evolved from pure noise, organized into layers, trained with virtual dopamine and pain and other proprietary methods — that generalizes to unseen data. Not a simulation of intelligence. A prototype of something new.

Demonstrated

Evolved from nothing

50+ independent evolutionary runs. Thousands of species emerged from random bytes. Three champions selected through competition. No human designed these neurons — nature did.

Demonstrated

Biological neuromodulation

Multiple neurotransmitter channels modulate learning, attention, and behavior — the same chemical vocabulary the brain uses. Reward and punishment aren't bolted on. They're structural.

Demonstrated

CPU-native — no GPU required

The entire architecture is designed for CPU cache hierarchies, not GPU tensor cores. Each layer fits in L1 cache (128 KB). Each functional region fits in L2 (4 MB). The brain grows by adding regions — limited only by available RAM. Irregular branching per unit, pointer chasing across layers, conditional tape execution — all things CPUs excel at and GPUs choke on. Runs on any device with a CPU. Learn more →

Demonstrated

Genuine generalization

Tested on data the circuit has never seen. Performance on unseen data matches or exceeds training accuracy — consistently, at every checkpoint. This is not memorization. The circuit learns transferable structure.

Demonstrated

Hot-swappable skill slots

Load specialized knowledge — medicine, law, coding — on demand. Bridge columns connect skills to the core circuit bidirectionally. LRU eviction preserves session learning. One deep skill can claim the entire budget.

Demonstrated

Neuroplasticity and myelination

The circuit actively prunes permanent connections that become wrong and replaces them with better ones. Connections proven by 8+ concurrent solves get myelinated — frozen insulated highways that carry signal at full strength. Self-correction without external intervention.

Demonstrated

Learns without backpropagation

No gradients. No loss functions. No optimizer states. Dopamine rewards connections that contribute. Pain weakens connections that harm. Evolution selects the computational species that survive. The circuit grows the way brains grow.

Demonstrated

Hippocampus and dreaming

A programmatic hippocampus records every experience — tagging each with reward, novelty, and confidence. During training it takes notes. During dream consolidation it replays the most important memories, strengthening verified pathways. The circuit learns while awake and consolidates while it sleeps — just like a brain.

Demonstrated

Human-readable vocabulary

250,000 real English words from GloVe 840B — not subword fragments. Every token is a word you can read. Contraction splitting, camelCase handling, and letter-by-letter fallback achieve 0% unknown rate on English text. 40 million tokens from 6 datasets tokenized and ready for training.

Early results

Eight parallel boxes — 12 layers each, 2,048 units per layer, 4 MB per functional region — trained from scratch on CPU. No pretrained weights. No transfer learning. 1,024-dimensional embeddings, 192 tokens of context, 80 tokens/second inference. Starting from random noise, reaching meaningful accuracy in seconds.

Trained data
52.0%
exact top-1 accuracy
69.5% top-10
out of 248,320 tokens
Unseen data
52.3%
exact top-1 accuracy
69.5% top-10
test exceeds train — genuine generalization
80 tok/s
standalone inference, no GPU
40M tokens
training pipeline from 6 datasets
4 MB
per functional region — fits in L2 cache

Trained and validated with train/test split on an Apple M4 Pro. 8 parallel boxes on 8 CPU cores. No GPU. No cluster. 40 million tokens of training data from 6 diverse datasets, tokenized with a 250,000-word human-readable vocabulary.

The circuit's proprietary verification mechanism identifies exactly which connections are responsible for correct predictions — something impossible in conventional neural networks where knowledge is distributed across billions of opaque weights. Only verified connections earn permanence. Everything else stays fluid, available for continued learning.

How fast is it?

Fast. A trained circuit runs faster than an untrained one — pruning and pain carve away weak connections, leaving sparser activation patterns that execute more efficiently. The circuit gets faster as it gets smarter.

At scale, the filing cabinet architecture means only active regions execute. A brain with thousands of functional regions but three active ones runs at the same speed as a three-region brain. Dormant knowledge costs nothing until it's needed.

Where it stands

A well-trained transformer LLM evaluated the same way — ranking the correct token against 1,000 random candidates — would score approximately 95–99% acc@1. The ENC at 52% has a real gap to close.

But consider what's being compared: a 4 MB circuit trained on CPU, versus models trained on trillions of tokens across thousands of GPUs over months. The ENC uses no backpropagation, no gradients, no optimizer states. It learns from scratch — from random noise to meaningful accuracy in seconds, to 52% with biological learning mechanisms alone.

Transformer LLM Evolved Neural Circuit
acc@1 (vs 1K candidates) ~95–99% 52%
Inference speed ~30–100 tok/s (consumer GPU) 80 tok/s, CPU-native, no GPU
Model size 4–70 GB Scales as needed
Training time Weeks on GPU clusters Hours on CPU
Context window 4K–128K tokens 192 tokens (8 parallel boxes × 24)
Training data Trillions of tokens 40M tokens
Learns after training No — weights can only be fine-tuned after creation Infere only or learn continuously, in real time

The question is not whether the ENC matches a transformer today. It does not. The question is whether this architecture — with its learning speed, storage efficiency, continuous adaptation, and a 40-million-token training pipeline ready to scale — can close the gap. The curve has room to climb.

What comes next

In Progress

Full-scale training

40 million tokens from 6 diverse datasets, tokenized with a 250,000-word human-readable vocabulary. Parallel functional regions — 8 boxes per region, each on its own CPU core — train with the full neuromodulation stack. The pipeline is built. Now it scales.

Near-term

Vision and audio

The circuit doesn't care what the input embeddings represent. Feed it image patches alongside text tokens and it learns cross-modal associations through the same biological mechanisms. One architecture for all senses.

Demonstrated

Two-tier memory

The brain handles short-term reasoning; confirmed prediction neuron traces are released to long-term memory (O(1) hash lookup, nanosecond retrieval). The brain stays plastic indefinitely — solved problems free capacity for new learning. The more it trains, the faster it gets.

On the horizon

Autonomous sleep cycles

The circuit detects when it's plateauing, saves its hippocampal journal, enters dream mode to consolidate, and wakes up better. Train-dream-train cycles that run autonomously. The programmatic hippocampus learns to become a neural one — routing by experience, not rules.

An AI that runs faster on CPU

Every other AI architecture is built on matrix multiplication — dense linear algebra that maps perfectly to GPU tensor cores. Without a GPU, they crawl. The Evolved Neural Circuit has no weight matrices. No matrix multiplies. No dense linear algebra at all.

Each computational unit runs a 16-byte tape program with branching, pointer chasing, and conditional logic. This is exactly what CPUs are built for — and exactly what GPUs are bad at. Thread divergence, random memory access, irregular control flow: these destroy GPU utilization but thrive on CPU cache hierarchies and branch prediction.

As small as 4 MB
per functional region — grows by adding regions
128 KB
per layer — fits in L1 cache
192
token context window per region

No matrix multiply

Zero GEMMs. Zero tensor core operations. The thing GPUs are specifically designed to accelerate doesn't exist in this architecture.

Irregular branching

Each of 2,048 units per layer runs its own tape with unique control flow. CPUs handle this per core. GPUs force lockstep execution — one divergent branch stalls the entire warp.

L1 cache sweet spot

Each layer is exactly L1 cache size on Apple Silicon (128 KB). The entire working set is always in the fastest memory on the chip. No cache misses. No memory stalls.

Parallel on CPU cores

8 independent boxes run on 8 CPU cores via GCD dispatch. Each processes its own context slice. No synchronization during forward pass. Clean scaling with core count.

A transformer needs a datacenter GPU and hundreds of watts to think. The Evolved Neural Circuit needs a cache line and milliwatts. The brain has no maximum size — it grows by adding functional regions, the way the human cerebral cortex is organized into 180 distinct areas per hemisphere. The hippocampus routes attention to the right regions and adds new ones as the brain learns. Each region is as small as 4 MB, fits in L2 cache, and runs on its own CPU core. The brain is limited only by available RAM — with mechanisms to swap specialized knowledge from disk when it exceeds that. It runs at full speed on a phone, a Raspberry Pi, a medical device, a satellite — anywhere there's a CPU. No cloud. No API call. No GPU allocation queue. The entire history of deep learning has been a story of scaling GPU compute. This architecture asks a different question: what if AI was designed for the processor that's already in every device on Earth?

A different kind of AI

Transformer LLMs Evolved Neural Circuit
Origin Human-designed architecture Evolved from random bytes
Training Backpropagation on GPU clusters Evolution + biological selection
Learns after deployment No — weights can only be fine-tuned after creation Continuously, in real time
Self-corrects No Neuroplasticity — prunes and rewires
Knows what it doesn't know No — guesses confidently Yes — uncertainty is structural
Hardware Datacenter GPUs, megawatts CPU-native, milliwatts
Modular expertise Monolithic — retrain everything Hot-swap skill slots on demand
Interpretable Black box — billions of opaque weights Glass box — every connection traceable
Multimodal by design Bolted on after the fact Architecture-agnostic input — any embedding