ExaBrain stores knowledge the way neurons might: every neuron has a name, every connection is a fact you can trace, and only the neurons that fire do any work. No training, no weights, nothing generated, so nothing made up.
Artificial neural networks compute densely: every neuron reads every one of its inputs, multiplies each by a weight, at every step. Simulate the human brain that way, with 86 billion neurons, 7,000 inputs each, updated a thousand times a second, and you need about 600 million billion operations every second.
The brain cannot be doing it that way. At any moment almost all of its neurons are silent, and a silent neuron costs nothing. The million-fold gap is an artefact of computing neurons that are not doing anything.
Even "mixture of experts" models still use tens of billions of weights for every word.
Around a million training steps, each adjusting every weight. A child learns a word from one example.
Knowledge is smeared across billions of weights: no source for an answer, no way to delete a fact.
Behaviour is steered from outside, with training and prompts, and guardrails can be talked around.
Opaque, ever more capable and only indirectly controlled: many of its builders now warn of catastrophe.
ExaBrain takes the brain's silence literally. It only ever touches the neurons that fire, and a firing neuron looks only at its own outputs. The cost of an answer depends on what is active, not on how much is stored.
Every neuron stands for one thing: a word pair in an abstract, an author, a chess position, a Bible verse. You can always ask what a neuron is.
A connection is either there or not: no weights to tune. Each one records something that was seen, so every answer can be traced back to its source.
New knowledge is stored by allocating neurons and connecting them, once. No gradient descent, no retraining, nothing old is overwritten.
Activity is a short list of the neurons that fire, never a vast array of numbers. Recall works by counting coincident inputs, which is cheap enough to run from an SSD.
Every demo runs entirely inside your web browser, on your own machine: nothing to install, and no server doing the work for you. The neurons, the updates per second and the recall scores you see are what your own computer achieves. Chrome or Firefox give the best results, because they let a demo use all of your processor's cores.
They are best taken in order: first what ExaBrain can do, then how conventional neural networks work, then the two side by side.
A cube of 2.25 quadrillion neurons, running in a browser tab. Huge numbers of neurons can be addressed because only the neurons that fire are processed: a silent neuron costs nothing.
Open the Cube →Millions of patterns, each learned from a single presentation and at very little cost. Give it a fragment and it returns the whole pattern. Labels can be attached to the patterns: show it an example it has never seen and 100,000 label neurons report how many stored examples it resembles. No training, no weights.
Open the Memory →A hands-on look at what happens inside a conventional neural network, built from floating-point weights and activation values: every unit connected to every other through a weight. At 1,024 × 1,024 units that is a trillion weights, the size of the largest language models, all used at every step.
Open FlopBrain →The same number of neurons, from 64 to 16 million, run as a dense network and as ExaBrain. Both get the same neurons and are charged for the storage and the work they really use. Watch how many updates per second each manages as the network grows, including with brain-like wiring of 7,000 inputs per neuron.
Open the comparison →A dense Hopfield-type memory and ExaBrain store the same sparse patterns among 4,096 neurons and recall them from a few cues. Run the capacity sweep as the number of stored patterns climbs into the millions: see where the dense memory stops recalling exactly, and how far beyond that ExaBrain goes.
Open the memory comparison →Every demo is a single page with nothing hidden: use your browser's View Source to read the code. Some may be simplified as the site develops.
Give it a set of cues and it returns every stored record that matches all of them, the way a database would, but with any combination of cues and no query planner. It searches the rarest cue first, so even a 100-cue question touches only a few neurons.
Try a query →Load a starting thought and let activity spread, a fixed number of neurons firing each cycle. Each cycle is remembered as an episode, so it knows when it has been somewhere before. First target: a bridge finder that wanders across hundreds of millions of research papers looking for analogies between fields that have never met.
Very different kinds of knowledge, stored the same way and searched together.
All of it sits on two OWC Express 4M2 Ultra RAID enclosures, 32 TB of SSD storage in all, attached to a single Mac Studio M3 Ultra with 256 GB of unified memory. No data centre, no cluster.
The search page works on a phone or a laptop. Type one cue per line, or press Random test query to see one of the million test questions answered. An access key is needed during the test phase: ask for one below.
curl -O https://exabrain.exaspike.com/exaload.py python3 exaload.py --url https://exabrain.exaspike.com \ --key [ACCESS KEY] --threads 8 --seconds 300
ExaBrain is built by Simon Thorpe, Directeur de Recherche Émérite at the CNRS in Toulouse and former director of CerCo, the Brain and Cognition Research Centre. His research has long focused on how the brain computes with spikes, from ultra-rapid visual processing to the idea that lifelong memories rest on simple, all-or-none synapses, the subject of his ERC Advanced Grant.
ExaSpike is the home of that work as it moves out of the lab.
For an access key, a demonstration, or to talk about research and partnerships.
simonjthorpe@mac.com