Mechanical Neural Network learns Addition through Gravity with pebbles (twitter.com)

Researchers built a physical "mechanical neural network" that learns to perform addition by routing pebbles under gravity. Inputs are encoded as physical configurations (e.g., tilted ramps or channels), and discrete pebbles act like trainable weights that alter how mass flows and settles; the resulting distribution is read out as the network’s output. The system is trained by adjusting pebble positions or counts to minimize addition errors, demonstrating that a simple, purely mechanical substrate can implement basic arithmetic through learning rather than hand-designed mechanisms. This work is significant because it broadens the notion of what constitutes a compute substrate for learning: instead of electrons and digital gates, computation emerges from classical mechanics, making the model inherently interpretable, low-power and robust to some noise sources. Key technical implications include analog, event-driven computation with discrete weight quanta (pebbles), an implicit nonlinearity from collisions and friction, and a training loop that maps parameter tweaks to macroscopic behavior. Limitations are clear—precision, speed, scalability and programmability lag electronic systems—but the demonstration points to new avenues for edge-class, energy-frugal neuromorphic devices, educational platforms, and hybrid digital-physical AI where physics itself becomes part of the model and optimization process.