Watch a Brain Learn

A digital “spark” agent with a four-neuron brain—two light-sensing neurons and two flipper-control neurons—can learn to steer toward light through trial and error, according to a new interactive demonstration. The brain’s wiring is reduced to three adjustable parameters: straight same-side connections, crossed opposite-side connections, and a drive setting that controls speed in darkness. Users manually tune these sliders to raise the spark’s score, defined as the average light encountered over 100 trials, with a target above 0.5. The demo then distills this manual tuning into a learning algorithm: before each trial the wiring changes randomly, and better outcomes reinforce the change while worse outcomes push wiring in the opposite direction. The work matters because it shows how reward and punishment can shape behavior in a minimal neural circuit, echoing how humans acquire skills like walking or biking. Next, watch whether the same algorithm scales to larger circuits and more complex behaviors beyond light-seeking.