Axon-2 / 01Sparse
Three percent awake.
Only the units that matter for a token fire. The rest stay silent, like cortex does.
≈ 3% active units per token
DWG 01Cortical sheetLive
We trace intelligenceone spike at a time.
Synapse Labs maps how real neurons pass signals, then builds models that learn the same way.
CAM +0.00 +0.00 +0.00
220 neurons · 0 axons · 0.0 Hz mean · cascade 00
Chapter 01 / 04
The network
Intelligence is traffic
01.1 — Signal flow
Every thought is a traffic pattern.
We record and simulate which paths light up in living cortex, how fast, and in what order. The pattern of traffic, not the size of the road, is what makes a brain clever.
01.2 — Scale
Big numbers, sparse traffic.
0bn
Neurons in a human brain
0tn
Synapses between them
0%
Firing at any moment
0W
Power for all of it
01.3 — Try it
What you are touching is a simulation.
Each neuron behind this page adds up its inputs, fires once it crosses threshold, then rests for half a second. Nothing is a canned loop — point at the cortex and it answers.
Chapter 02 / 04
One neuron
N-001, pyramidal, layer 2/3
02 — Anatomy of a decision
Ten thousand inputs. One answer.
Dendrites listen
Branches collect thousands of small nudges from other cells.
The soma decides
If the sum crosses about −55 mV, the cell commits. There is no half spike.
The axon speaks
A pulse runs down the axon at metres per second and splits at every branch.
Chapter 03 / 04
The gap
Where learning actually lives
03 — The synapse
Learning is a change in how loudly one cell speaks to another.
When the pulse reaches the end of the axon, vesicles spill messenger molecules across a gap narrower than a virus. How many cross, and how well they are heard, is the weight. We study that gap, and we build it into safety: every weight Axon-2 changes is logged and reversible.
0nm
Width of the cleft
0ms
Time to cross it
0
Number that learns
Chapter 04 / 04
The thought
Seven layers, same rules
04 — Axon-2
A model that fires only when it has something to say.
Axon-2 / 02Online
Learns at the gap.
Weights update during use, inside limits you set, instead of waiting for a nightly retrain.
Bounded plasticity · per-tenant
Axon-2 / 03Traceable
Shows its work.
Every answer ships with its activation trace: which layers fired, in what order, and why.
Trace export · JSON / Parquet
Axon-2 / 04Frugal
Runs on a desk.
Event-driven inference means the hardware idles between spikes. One workstation serves a lab.
Single-GPU inference
Axon-2 / 05Open
Weights for research.
Universities get the full weights and the simulator you just played with, under a research licence.
Research licence · 2026
Research access · Cohort 04
Put your hand on the network.
We open Axon-2 to twelve labs per cohort. Tell us what you want to trace, and we will send the weights, the simulator and a person who built them.