RL is more information inefficient than you thought
https://www.dwarkesh.com/p/bits-per-sample
#HackerNews #RL #information #inefficiency #AI #research #machinelearning #dataanalysis
🧠 New preprint by Codol et al. (2025): Brain-like #NeuralDynamics for #behavioral control develop through #ReinforcementLearning. They show that only #RL, not #SupervisedLearning, yields neural activity geometries & dynamics matching monkey #MotorCortex recordings. RL-trained #RNNs operate at the edge of #chaos, reproduce adaptive reorganization under #visuomotor rotation, and require realistic limb #biomechanics to achieve brain-like control.
