Vicarious

Contribution to AI

Deep learning's success in vision came with an unpaid bill: millions of labelled examples for tasks a person picks up from a handful. Vicarious spent a decade arguing that the bill could be settled with structure rather than data, and produced the sharpest published evidence anyone offered for that position. The 2017 Science paper described a hierarchical generative model of contours and surfaces that broke text CAPTCHAs — 66.6 per cent on reCAPTCHA, against the one per cent threshold at which such a test is considered defeated — after training on a few clean examples of each character rather than millions of distorted ones, some three hundred times less data than convolutional networks required on comparable scene-text benchmarks. Because the model was generative, classification, segmentation and reasoning about what occluded what were queries on a single representation instead of separate systems bolted together. Schema Networks made the equivalent case for control: an object-centred causal model learned from Breakout that kept scoring when the paddle was offset or the bricks rearranged, exactly the variations under which the deep reinforcement learners of the day fell apart. Neither result displaced the methods it criticised, and the later years went into industrial picking sold by the unit. What survives is the counter-example, and a research group that carried the argument into DeepMind when the company was split up in 2022.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).