Minsky’s 1961 argument is visualized here as a clash between two stories: reward can strengthen a path, but only a good representation can make nearby situations count as meaningfully similar. The page stays diagram-first: random topology, similarity geometry, transfer-vs-interference, and the later lineage from perceptrons to random features and echo state networks.
The overview diagram separates four learning stories from the transcript: rote memory, associative recall, pattern recognition, and true generalization. Hover any block or connection to see what the random-net proposal can plausibly do and where it fails.
The heatmap and scatter view show the paper’s central objection. Reward can mark a successful response after the fact, but if the hidden geometry does not cluster related cases, spillover becomes interference rather than generalization.
This lab compares three mock regimes: pure internal reinforcement in a random tangle, fixed random features with a trained readout, and a learned representation. Toggle to see which setup converts scale into useful separation instead of expensive ambiguity.
The timeline maps what survived. Randomness persisted, but only after later work isolated the random part from the trainable part, or wrapped it inside a precise objective, memory system, or optimization rule.