In 1948, Claude Shannon laid out a definition of information that still holds up: information is surprise. Not data volume. Not signal strength. Surprise — the gap between what you expected and what actually happened.

That idea has a direct, practical consequence for building an energy-efficient AI architecture: you should only spend energy when something unexpected occurs. The rest of the time, a system should be able to coast — make its predictions, keep running, and save the real energy spend for the moments that actually require correction.

Most AI models today, including the large neural networks behind modern AI, don’t operate on that principle. Every time a neural network makes a prediction, it touches every single parameter in the model — regardless of whether anything surprising is happening. And during training, when the model corrects itself, it touches every parameter again. Sometimes more than once.

That’s a lot of energy spent on business as usual. The architecture doesn’t distinguish between “everything is going as expected” and “something just happened that requires a real update.” It treats every moment the same.

Brain-CA’s Estimator takes the opposite approach. Inference — making a prediction — is designed to be cheap, low-energy work. It’s only when a prediction turns out to be wrong that the system spends energy to correct itself. Prediction is the default state. Correction is the exception, and it’s treated like one.

This connects to two ideas that guide how we think about efficient AI:

• Learning is compression. A system observes a large amount of information and compresses it down, keeping what mattered — the surprises — and discarding the rest. Efficient learning isn’t about storing everything; it’s about storing what was unexpected.
• Inference is routing. When it’s time to make a decision, the goal is to involve the fewest parameters necessary — not the whole model. A decision tree gets you there: it routes you down a path to the relevant parameters directly, rather than touching everything to arrive at an answer. Neural networks aren’t organized this way. They hit every parameter, every time, whether it’s needed or not.

This is more than an academic distinction. It’s the difference between an energy-efficient AI architecture that scales its energy use with actual computational need, and one that pays the same energy cost regardless of what’s happening. As AI inference becomes the dominant cost driver in AI infrastructure, that difference compounds fast.

It’s also, fundamentally, the Brain-CA advantage: an energy-efficient AI architecture designed from the ground up around a 75-year-old information theory principle that the rest of the industry built around a different set of assumptions.

Shannon gave us the theory. Here’s how we built the architecture around it.