Intelligence built on the brain's principles. Learns for life, never forgets.

Stiefel and Coggan, The energy challenges of artificial superintelligence (Frontiers in Artificial Intelligence, 2023), argue that contemporary semiconductor computing poses "a significant if not insurmountable barrier" to intelligence beyond our own, because such a system would be "orders of magnitude less efficient in energy use than human brains" and would consume more power than a highly industrialised nation generates. Their measurement: simulating ten million cortical neurons for one second of biological time took 270 kilowatts and eight hours, thirty thousand times slower than nature, which puts biological computing at least 9 × 10⁸ times more energy-efficient than the silicon running today's AI. Their diagnosis is that current AI is "uni-scale" and misses the brain's multi-level architecture. Their prescription is biomimicry: discover the "bioflop" and the "bioalgorithms" and engineer a manageable equivalent. Estirion is an attempt at exactly that on today's hardware. We do not claim to have closed the gap; we claim to be building on the right principles, and we measure as we go.

Where. Every neuron has a place. Who it can talk to depends on distance. Signals. In sparse, event-driven pulses, on or off. A silent neuron costs nothing, one of the reasons brains are cheap to run. Wires. By use; neurons that fire together grow closer. Grows. A new word grows a neuron. Nothing is set aside in advance. Learns. Locally, neuron by neuron, with a small conventional learner on top where it helps, protected so nothing already learned is lost. Answers. By which neurons fire and in what order.

Space does the work. Neurons have real positions, so distance decides who talks to whom, the way geography decides who meets. Copy the brain's rules, not its look. Pulses, areas that specialise on their own, growth where there is demand, memories hardened the way sleep hardens the day's learning. Many ideas share the same neurons. Patterns overlap in one space, which is how a hundred-thousand-word vocabulary grows in under a minute. Structure instead of size. Memory instead of re-reading an ever-longer transcript, growing instead of retraining. Correct first, fast second. A slow reference run defines correct; the fast version must match it to the last digit. How each is realised is the protected part, under NDA until the patent is filed.

Every figure here is one the live system itself produces; anyone with the demo can reproduce it. The 0.275× is an operation count, not yet a measurement in joules. Converting sparse operations into wall-clock and energy savings on real silicon is the engineering still ahead, and it is named as such on the risk slide. The forgetting test is Estirion against itself with one mechanism off. No other company's model was run, so it is a clean test of the mechanism, not a race.

Is: a small, honest assistant that improves. Holds a real conversation, follows instructions, shows its reasoning, learns live without erasing the old. Is not: as broad or fluent as the biggest models, and we will not claim otherwise. Honest limit: learning without forgetting is measured today. That it compounds into a better assistant, lesson by lesson, is not yet shown, and is exactly what v1 exists to demonstrate.

Two separate risks, and we put both in front of every investor before they ask. First, whether learning keeps scaling with the no-forgetting property intact. Second, whether sparse operations become a real energy advantage over dense hardware at scale. The compute and the round exist to answer them.