University of California, Berkeley

Contribution to AI

A set whose members belong to it only partly was an odd thing to propose in 1965, and Lotfi Zadeh's fuzzy sets paper was received coolly by an American field then committed to crisp symbols. Control engineers took it seriously instead, and by 1987 a fuzzy controller was driving trains on the Sendai subway, followed by a decade of Japanese appliances tuned on linguistic rules rather than differential equations — the first large demonstration that vague categories could be engineered. Thirty years on, Berkeley's second contribution was pedagogical: Russell and Norvig's 1995 textbook reorganised the subject around the rational agent acting under uncertainty, so probability and decision theory became the spine of an introductory course rather than an advanced topic, and more than a thousand institutions now teach from a shared vocabulary. The deep learning era brought infrastructure. DeCAF and then Caffe, released from the vision group in 2013 and 2014, made pretrained weights something you downloaded and fine-tuned instead of retrained, and the Model Zoo established the habit of publishing weights years before anyone built a business on it. The robot learning work that followed produced trust region policy optimisation and soft actor-critic, algorithms other laboratories still use as baselines, with Ray and RLlib supplying the distributed machinery underneath. Human-Compatible AI, founded in 2016, gave alignment research an academic home and a technical framing: uncertainty about the objective, not a fixed one.

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