AI from zero
From Turing's imitation game to the chatbot on your phone: how machines learned to learn, why AI froze twice, and the risks it carries today.
8 lessons · 52 min
- 1Turing and the imitation gameFreeIn 1950 Alan Turing swapped an unanswerable question for a test you could actually run. Seventy-odd years later we still argue about it.1950 · 6 min
- 2Dartmouth 1956: the field gets a nameFour researchers asked for a summer to study thinking machines and gave the field its name. The first approach they backed: symbols, rules and search.1955–1956 · 6 min
- 3The perceptron and the first AI winterIn 1958 a machine learned to tell left from right and the press promised thinking computers. What it could do, what it could not, and the first winter.1958–1973 · 7 min
- 4Expert systems and the second winterIn the 1980s AI went to work as rule books typed in by experts. It paid off, then hit a wall: knowledge is expensive to write down and brittle to keep.1972–1992 · 6 min
- 5How a neural network learnsWeights, a loss, a gradient and a small step downhill, repeated millions of times. The training loop behind every modern model, without the maths.1986 · 7 min
- 6ImageNet and AlexNet: data meets GPUsIn 2012 a neural network trained on two gaming graphics cards cut the error rate of the best image recognisers almost in half. Deep learning had arrived.2006–2012 · 6 min
- 7Transformers and how a language model writesA 2017 paper threw out recurrence and kept attention. Scaled up, that design became the large language model, which writes one token at a time.2017–2020 · 7 min
- 8AI risks and security todayModels that invent facts, obey hidden instructions, learn from poisoned data and inherit bias. The main risks, the defences, and the EU's AI Act.2018–today · 7 min