Introduction
In 1986, a paper changed the trajectory of artificial intelligence. It was called "Learning Internal Representations by Error Propagation," and it was co-authored by a small group of researchers including a young psychologist named David Rumelhart and a cognitive scientist named Geoffrey Hinton. The paper described a simple method for training multi-layer neural networks: you feed data forward through the network, compute the error at the output, and then propagate that error backward through each layer, adjusting every connection just a little bit in the direction that reduces the error.
They called it backpropagation — backprop for short — and it was the spark that lit the AI revolution.
Nearly forty years later, backprop is everywhere. It is the learning algorithm behind every large language model, every image generator, every recommendation system, every autonomous vehicle. The modern AI industry — worth hundreds of billions of dollars — runs on a single mathematical trick invented in the 1980s. It is, by any measure, the most successful learning algorithm ever devised.
There is just one problem. Backprop cannot be what the brain does.
The objection is not philosophical. It is not a matter of "the brain is too complex to understand." It is a specific, mechanical argument with three distinct components — three implausibilities that make backprop an impossible account of biological learning. Neuroscientists have known about them for decades. AI researchers have largely ignored them, because backprop works so well that the biological question seemed like someone else's problem.
But the biological question is coming home to roost. Because the human brain learns with about 20 watts of power — less than a dim light bulb. The largest AI models consume megawatts. The brain processes information asynchronously, event-driven, with no global clock. It learns from a handful of examples. It does not forget what it learned yesterday when it learns something new today.
The gap between AI and the brain is not a gap in capability — in narrow domains, AI has already surpassed us. The gap is in efficiency. And that efficiency gap traces back, in large part, to the fact that the brain is not running backprop.
This book is about the search for a learning algorithm the brain could actually use — and what that search tells us about where AI is heading. The story runs through neuroscience, physics, and computer architecture, and it has an unexpected protagonist: a small startup called Zyphra that is building a different kind of inference engine.
If they are right, the future of AI does not look like bigger GPUs running faster backprop. It looks like analog hardware running algorithms that learn the same way neurons do — by being nudged by the world and adjusting accordingly.