Soverand · Brain-inspired efficient AI
Intelligence
density.
More intelligence. Less data, less compute, less energy.
Today, artificial intelligence depends on ever more data, more compute and more energy. That path has concentrated it in very few hands. We research how to break that dependence, so it can reach anywhere and anyone.
01 / The limits of scale
Today's AI advances by sheer scale.
Each new generation of models multiplies the resources it needs, and every improvement costs more than the last. The path works, but only a few can afford to follow it.
- Data
It needs vast numbers of examples to learn anything new.
- Frozen knowledge
What it learns is frozen once training ends. Updating it means training again.
- Cost
The compute and energy it requires grow with every new generation of models.
- Text only
It knows the world through what has been written about it, not through how it works.
The next leap won't come from scale alone.
It will come from efficiency.
02 / The proof
The proof is in your head.
Your brain learns continuously, from a handful of experiences, on about 20 watts. Training today's frontier AI takes campuses that draw over a hundred megawatts. The gap is not physics. It is architecture.
Since you opened this page
Estimates: about 20 W for a human brain, about 150 MW for a large AI training campus.
Inspired by the brain
What we take from the brain.
We don't copy its biology. We take the principles that make it so efficient, and turn them into architecture.
- Hippocampus
Fast memory. Each experience is stored the moment it happens.
- Sleep
Consolidation. Experience becomes lasting knowledge without overwriting what came before.
- Sparse firing
Sparsity. Only a small fraction of neurons fire at any moment.
- Prediction
World models. The brain constantly predicts what comes next, and learns from its mistakes.
03 / Our focus
Intelligence density.
How much intelligence a system gets from every resource it consumes. It is the measure we hold everything we build to.
- Less data
Learns from a few examples, not millions.
- Less compute
Takes in what it learns without retraining.
- Less energy
Designed to run on the power of a personal computer, not a data center.
04 / What we build
AI that keeps learning.
A model stops learning the day its training ends. Our architectures keep learning as they are used, the way a brain does, and work alongside existing language models instead of replacing them.
Few examples
A handful of cases is enough to learn a new task.
Continual learning
Every use makes it a little better.
No forgetting
Learning something new doesn't erase what it already knew.
Tailored
Learns from the data and experience of each person or organization, without that data ever leaving their own machines.
05 / Understanding the world
Beyond text.
Language models know the world through what has been written about it. They can describe a glass falling, but they don't understand why it breaks. We research systems that learn from the world itself: they observe, anticipate what will happen, and correct themselves when they are wrong.
- Observe
Learn from images, video and sensors, not just words.
- Anticipate
Predict what will happen, not just describe what already has.
- Correct
Compare what was expected with what actually happens, and learn from the difference.
Ongoing research
06 / From algorithm to hardware
Efficiency at every layer.
Our architectures run on today's hardware. But we design them to take advantage of the chips to come, which will be far more efficient than today's.
- Sparsity
Event-driven chips only spend energy on the units that fire. Our architectures keep that number small.
- Local learning
Each connection learns from what happens around it, without propagating errors through the whole network. It is the kind of learning neuromorphic chips make possible.
- Noise tolerance
Analog and photonic chips use far less energy, but make small errors in every calculation. Today's models can't cope with them. We design architectures that work despite that noise.
- Beyond the GPU
LLMs rely on operations that are only efficient on GPUs, such as attention over long contexts. Our architectures don't.
- New hardware
We are exploring how to bring these architectures to neuromorphic, photonic, analog and in-memory computing chips.
New hardware: ongoing research
07 / Method
Measured, not promised.
We compare every result against the best current models, under the same conditions. And every experiment answers four questions.
- How many examples does it need?To learn something new, and learn it well.
- How much does it forget?Of what it already knew, when it learns something new.
- How much does it cost?In compute, memory and energy, both to answer and to learn.
- Is it ready for the hardware to come?How much of it is active at once, and how much noise it can tolerate.
The brain is the proof.
Efficiency is the path.
Contact
Want to know more about what we're building? Get in touch.