hawkes.aiKey ideas
- The one idea — The perceptron (Rosenblatt, 1958) was the first machine that learned from examples — but Minsky & Papert's 1969 proof that a single layer can't learn XOR, alongside funding cuts, froze the field into the first AI winter; the surviving idea was connectionism.
- How it is shown — A single glass neuron learns, the press inflates it into impossible promises, a logic-plane proves one layer can't slice XOR, and the whole scene freezes over — with one warm ember (connectionism) refusing to die.
- The trap to avoid — Hearing a narrow, true proof (one-layer perceptrons can't do XOR) as a broad, false verdict (neural networks are a dead end) — the loophole was always a hidden layer.
Before transformers, before GPUs, before any of this worked — there was a machine that learned from examples. Nineteen fifty-eight.
The one idea
The perceptron (Rosenblatt, 1958) was the first machine that learned from examples — but Minsky & Papert's 1969 proof that a single layer can't learn XOR, alongside funding cuts, froze the field into the first AI winter; the surviving idea was connectionism.
It made headlines, then it made a winter. This is where our story actually begins. First, the idea. In nineteen forty-three, McCulloch and Pitts wrote the neuron as math: inputs, weights, a threshold, an output. Then Frank Rosenblatt built it. His perceptron — a learning machine at the Cornell Aeronautical Lab, soon wired to a grid of light sensors — didn't just compute. It adjusted its own weights until it classified patterns correctly. It learned. A machine that learned from examples, in nineteen fifty-eight — the press lost its mind. The New York Times reported the Navy expected a machine that would walk, talk, see, write, and be conscious of its existence. Rosenblatt himself made bold predictions. The optimism ran years ahead of the mathematics.
How it works — the demo
A single glass neuron learns, the press inflates it into impossible promises, a logic-plane proves one layer can't slice XOR, and the whole scene freezes over — with one warm ember (connectionism) refusing to die.
Then came the reckoning. In nineteen sixty-nine, Marvin Minsky and Seymour Papert published a book called Perceptrons. With careful proofs, they showed a single-layer perceptron cannot learn even the simplest logic — exclusive-or, XOR, deciding whether two inputs differ. No line separates that problem. One layer of neurons, they proved, was fundamentally blind to it. The timing was brutal. Governments were already losing patience. In nineteen seventy-three, Britain's Lighthill report judged AI's grand promises unmet, and UK funding collapsed. American money tightened too. Between the proofs and the budgets, neural networks went cold. Historians call it the first AI winter — and it lasted years. Here's the trap.
The trap to avoid
Hearing a narrow, true proof (one-layer perceptrons can't do XOR) as a broad, false verdict (neural networks are a dead end) — the loophole was always a hidden layer.
Why it matters — and what’s next
Minsky and Papert proved something true — about single-layer networks. The field heard something false: that neural networks were dead. But the proof had a loophole. Add a hidden layer, and the XOR wall falls — nobody yet knew how to train it. The idea survived: connectionism, intelligence from networks of simple units. So the neuron went dormant, but not dead. The perceptron's failure wasn't the idea — it was a missing algorithm, one that could teach a hidden layer. That algorithm existed. It just hadn't been connected to this problem yet. When it was, the second act began. Next: backpropagation.
This is a supplement in AI: Zero → Frontier — a side-trip that deepens the act it sits beside, one file and one loop at a time.
Full transcript 2:45 of narration
Before transformers, before GPUs, before any of this worked — there was a machine that learned from examples. Nineteen fifty-eight. It made headlines, then it made a winter.
This is where our story actually begins. First, the idea. In nineteen forty-three, McCulloch and Pitts wrote the neuron as math: inputs, weights, a threshold, an output.
Then Frank Rosenblatt built it. His perceptron — a learning machine at the Cornell Aeronautical Lab, soon wired to a grid of light sensors — didn't just compute. It adjusted its own weights until it classified patterns correctly.
It learned. A machine that learned from examples, in nineteen fifty-eight — the press lost its mind. The New York Times reported the Navy expected a machine that would walk, talk, see, write, and be conscious of its existence.
Rosenblatt himself made bold predictions. The optimism ran years ahead of the mathematics. Then came the reckoning.
In nineteen sixty-nine, Marvin Minsky and Seymour Papert published a book called Perceptrons. With careful proofs, they showed a single-layer perceptron cannot learn even the simplest logic — exclusive-or, XOR, deciding whether two inputs differ. No line separates that problem.
One layer of neurons, they proved, was fundamentally blind to it. The timing was brutal. Governments were already losing patience.
In nineteen seventy-three, Britain's Lighthill report judged AI's grand promises unmet, and UK funding collapsed. American money tightened too. Between the proofs and the budgets, neural networks went cold.
Historians call it the first AI winter — and it lasted years. Here's the trap. Minsky and Papert proved something true — about single-layer networks.
The field heard something false: that neural networks were dead. But the proof had a loophole. Add a hidden layer, and the XOR wall falls — nobody yet knew how to train it.
The idea survived: connectionism, intelligence from networks of simple units. So the neuron went dormant, but not dead. The perceptron's failure wasn't the idea — it was a missing algorithm, one that could teach a hidden layer.
That algorithm existed. It just hadn't been connected to this problem yet. When it was, the second act began.
Next: backpropagation.