Research
Early-stage · Unreleased

We think intelligence shouldn't need
a million labeled examples.

Every product we ship today — Evi, Kus, Revmosys — runs on the same kind of AI everyone else builds on. In parallel, we're researching something fundamentally different: a learning architecture modeled on how brains actually work, not on scaling the current approach further.

The starting point

Today's AI is extraordinary — and built on one specific idea.

Large language models, including the ones behind our own products, learn by comparing a guess against a correct answer millions of times and adjusting the whole network backward from that error — backpropagation. It needs enormous datasets and enormous compute, and once training stops, the model is frozen: it doesn't keep learning from what happens after it ships.

The bet

An architecture modeled on the brain, not the transformer.

We're building a substrate of spiking neurons that learn through local rules — connections strengthen between neurons that activate together, the same principle real synapses use — rather than one global error signal reaching backward through an entire network. There is no backpropagation anywhere in this core. It's domain-agnostic: the same substrate that learns to organize written problems is the one we're testing on other kinds of sequential data.

We're not publishing the specifics of what we've found yet — that comes later, as real results earn it. This page is about the direction, not a method disclosure.

How it differs

Not a claim that we've beaten anything. A different bet on how learning should work.

Today's AI (what we ship on)
Our research bet
Learns from massive labeled or token-prediction datasets, at huge compute cost.
Learns from raw exposure to real data — no labeled-answer comparison required.
A single global signal (backpropagation) adjusts the whole network at once.
Local learning rules adjust connections directly — closer to how real synapses change.
Training and deployment are separate: the model is frozen once it ships.
Designed to keep learning from what it encounters, not just during a training run.
Proven at scale — this is why it powers every major assistant today, including ours.
Early-stage and unproven at scale. That's exactly what we're now testing.
Where it actually stands

What's real today, plainly stated.

This is active research, not a shipped product and not a benchmark result. Today the system can take in real data and organize what it has seen. It cannot yet solve new problems, generate answers, or do anything user-facing — and we're not going to claim otherwise to make the story sound further along than it is.

Where we think this goes

A long-term bet, said plainly as one.

We don't think scaling today's approach further is the only path to more general intelligence. Our bet is that a system which keeps learning from raw experience, the way brains do, instead of stopping once training ends, is a necessary ingredient for something closer to general intelligence — not just a bigger version of the chatbots that exist today. That's speculative, it's early, and it may not work. We'd rather say that directly than dress up early research as a finished result.

Following this closely?

We'll share real results here as they happen — not before. For anything else, write to us directly.