A Brief History of AI

Why history matters here

AI is a field of recurring manias and hangovers. Knowing the cycle inoculates you against both hype and dismissal — and explains why today's moment is genuinely different in some ways and familiar in others.

1950s–60s: birth and first euphoria

Alan Turing asks "can machines think?" (1950). The field is christened at the Dartmouth workshop (1956), whose proposal predicted major progress in one summer. Early wins dazzle: programs prove theorems, solve algebra word problems, play checkers. Frank Rosenblatt's perceptron (1958) — a learning artificial neuron, ancestor of everything in your DL path — makes the New York Times, which relays predictions of walking, talking machines. Pioneers forecast human-level AI within a generation.

1970s: the first winter

Reality bites. Translation flops, robots can't cross a room, and Minsky & Papert (1969) prove single-layer perceptrons can't even learn XOR (you know the fix — hidden layers — but training them wasn't practical yet). Funding agencies read damning reports (Lighthill, 1973) and pull out. Lesson one of AI history: toy-problem success does not scale for free — the "combinatorial explosion" of the real world devours naive methods.

1980s: expert systems boom… and second winter

Commercial AI returns as expert systems: thousands of hand-written IF-THEN rules encoding specialists' knowledge (XCON saved DEC millions configuring computers). A billion-dollar industry blooms — then wilts: rule bases prove brittle (no common sense, absurd edge-case behavior) and unmaintainable (updating ten thousand interacting rules by hand). By 1987 the market collapses; "AI" becomes a dirty word in grant applications for a decade. Meanwhile, quietly, backpropagation is popularized (1986) — the seed of everything to come.

1990s–2000s: statistics wins quietly

AI rebrands as machine learning: probabilistic methods, SVMs, ensembles — modest claims, measurable progress, real deployments (spam filters, fraud detection, search ranking). Deep Blue beats Kasparov at chess (1997): a triumph of search + hand-tuned evaluation. The web begins generating the giant datasets the next act will need.

2012–present: the deep learning era

The ingredients finally align — big data (ImageNet), parallel compute (GPUs), algorithmic touches (ReLU, dropout): AlexNet (2012) shatters vision benchmarks and the field pivots overnight. Milestones cascade: AlphaGo defeats Lee Sedol (2016) at a game thought a decade away; transformers (2017) unlock scaling; AlphaFold (2020) cracks protein-structure prediction, a 50-year grand challenge; ChatGPT (2022) puts conversational AI in hundreds of millions of hands and makes "generative AI" a household phrase.

The pattern to carry forward: capabilities arrive in sudden jumps when data, compute and algorithms align — and each boom's rhetoric overshoots before the next consolidation. Judge claims against those three ingredients, not against headlines.

Key takeaways

• Two AI winters followed overpromising: scaling walls (70s) and brittle rules (80s).
• The statistical turn traded grand claims for measurable progress — and won.
• Deep learning's rise = data + compute + algorithms aligning; that trio remains the lens for evaluating every new claim.