Conclusion
The search for a biologically plausible alternative to backpropagation is not an academic exercise. It is a practical engineering problem with billion-dollar implications.
The argument is simple. The brain learns with approximately 20 watts of power. The largest AI models consume millions of times more energy to learn tasks that a human child could master in minutes. If we could close even a fraction of that efficiency gap, the implications would be transformative: AI that runs on edge devices, AI that learns continuously without forgetting, AI that does not require datacenters with their own power plants.
The path to that efficiency runs through biologically plausible learning. Not because biology is sacred, but because biology is the only existence proof we have that efficient learning is possible. The brain is the only system we know of that can learn complex tasks from limited data using limited energy. Understanding how it does that is not neuroscience for its own sake. It is reverse-engineering the most efficient learning machine in the known universe.
The alternatives to backprop — equilibrium propagation, the forward-forward algorithm, predictive coding — are not ready to replace it. They are less accurate, slower, and less well understood. But they share a set of properties that backprop lacks: local learning rules, no requirement for weight symmetry, no global error signal, no separation between inference and learning. These properties are not optional. They are the necessary conditions for any algorithm that could run on the hardware we are building.
The bet that companies like Zyphra are making is that the algorithm and the hardware must be designed together. The model architecture, the learning rule, and the physical substrate are not independent variables. They are coupled. A model designed for a diffusion engine requires a different learning algorithm than a model designed for autoregressive decoding. A learning algorithm designed for local, asynchronous hardware requires a different model architecture than backprop does.
The convergence of these three threads — brain-inspired algorithms, neuromorphic hardware, and co-designed model architectures — is the most interesting development in AI today. It is less visible than the race to build bigger models, but it addresses a more fundamental question: not whether we can build AI that is more capable, but whether we can build AI that is more efficient.
If the answer is yes — and there is no reason to think it is not — then the AI of the future will not look like a bigger version of what we have today. It will look fundamentally different. It will learn the way the brain learns. It will compute the way the brain computes. And it will consume energy the way the brain consumes energy.
That future is not here yet. But the path to it is visible, and the first steps are being taken.