From McCarthy & Rosenblatt to agentic AI — the lineages, people, and paradigm shifts that produced ChatGPT, Claude, Gemini and more

Deterministic — Codified Rules: "Same input → same output, every time"

Probabilistic — Emergent Patterns: "Most likely output, learned from data"

Era 1 · 1958–1985 — Two Parents & Symbolic Dominance

Founders

• John McCarthy — LISP (1958). Codified rules, symbolic logic, explicit if-then reasoning
• Frank Rosenblatt — Perceptron (1958). Weighted connections learn from data. The probabilistic seed

Symbolic AI Wins

• 1969 — Minsky & Papert publish Perceptrons — mathematically "kill" neural nets. Funding collapses
• 1970s–80s — Expert Systems dominate: DENDRAL, MYCIN. Hand-coded rules, brittle in practice
• Early agent concepts: cybernetics, feedback loops, BDI (belief-desire-intention) model
• 1980s — Rodney Brooks — subsumption architecture. Reactive, modular robots. Precursor to modern agent design

Era 2 · 1986–2009 — Deep Learning & Trial-and-Error

Godfathers

• Geoffrey Hinton (Toronto), Yann LeCun (NYU), Yoshua Bengio (Montréal) — kept connectionism alive through decades of indifference
• 1986 — Backpropagation paper (Rumelhart, Hinton, Williams). Shows how to train multi-layer neural networks

Reinforcement Learning

• 1988 — Richard Sutton — temporal difference learning. Agents learn by trial and reward