Inference Is Learning: No Phase Separation in River Learning

Modern deep learning draws a strict boundary between training and inference. During training, a for ward pass is followed by a backward pass and a weight update; during inference, weights are frozen and only the forward pass is executed. This separation is ac cepted as a foundational assumption, yet it has notheoretical necessity. We show that River Learning [1] operates without this separation. A single carve() operation simultaneously performs prediction (inference) and resistance update learning). Three properties emerge from this structural fact alone: Unsupervised learning: semantic structure is discovered without any labels. Interpretability: each node directly explains what it learned. Continual learning: weights update during inference, with no mode switch. In an unsupervised word representation experiment at Word2Vec scale: WS353 Spearman correlation of 0.2576 is achieved (vs. Word2Vec dim=300 ≈ 0.30), with 1.6M parameters (27% of Word2Vec dim=300), no negative sampling, no labels, and no gradient. Eighty highway nodes self-organize into interpretable semantic clusters without any supervision.