Source: Digital Wetware — connectomics and zero-shot biological AI
The term “zero-shot wetware” might be the most evocative phrase I’ve encountered in AI research. It describes an approach to building intelligent systems that is almost the opposite of how current AI is built: instead of training a model from scratch on large datasets, you map the structural wiring of a biological organism and run it digitally. Zero training. Zero scripting. Zero reinforcement learning. Just structure, and electricity.
The contrast with deep learning is stark. Standard neural network training starts with random weights and adjusts them through backpropagation over billions of examples. Connectomics-based AI starts with millions of years of evolutionary pre-training encoded in physical synaptic architecture.
The FlyWire connectome — 125,000 nodes and 50 million routing intersections in the Drosophila brain — is the most complete example we have. When this map is translated into a computational substrate and the neurons are activated, emergent behaviours appear: walking, grooming, foraging. Not because these behaviours were programmed or trained, but because the structural wiring that produces them in the biological organism produces them in the digital replica too.
The implications: if structural priors can encode sophisticated behaviour without training data, then the most efficient path to certain kinds of intelligence might not run through scale at all — it might run through better understanding of biological structure.