A new AI system called 'noids' replaces traditional boid rules with a neural network that learns flocking behavior from data. Instead of hand-coded rules, each agent tracks 5 nearest neighbors and uses 1,922 learned parameters to decide steering. The network takes 24 inputs - velocity, heading, and relative positions of neighbors - and outputs acceleration forces.

This approach builds on real starling research showing birds track ~7 topological neighbors rather than using global awareness. The neural network uses three layers with SiLU activations to transform local perception into steering decisions. Training happens through imitation learning, where the network copies classic Reynolds rules' behavior. The entire behavioral model fits in a QR code-sized parameter set.

The system demonstrates how complex group behavior emerges from simple local interactions. By learning from biological data rather than programming explicit rules, noids achieve realistic flocking without requiring designers to tune weights or invent mathematical formulations. The approach could extend to other multi-agent systems where local perception drives collective intelligence.