History of AI — Tutorial

The first 70 years: logic, rules, and two winters

AI has a longer history than most people realise. Alan Turing asked "Can machines think?" in 1950 and designed the Turing Test as a measure of machine intelligence. The field was formally founded at the Dartmouth Conference in 1956, where researchers predicted human-level AI within a generation.

The 1960s and 70s produced early successes: ELIZA (1966) simulated a therapist using pattern matching. SHRDLU (1970) could manipulate blocks in a virtual world. But these systems were brittle — they only worked in tiny, controlled domains.

The first AI winter (1974–1980) arrived when governments cut funding after promised results didn't materialise. The second winter (1987–1993) followed the collapse of expert systems — rule-based programs that required humans to manually encode knowledge. Both winters shared the same root cause: researchers underestimated how hard general intelligence is.

The neural network revival and the deep learning explosion

The modern era began with three pivotal moments:

1986 — Backpropagation: Rumelhart, Hinton, and Williams showed that multi-layer neural networks could be trained efficiently using the chain rule. This had been known theoretically but was now practical.

1998 — LeNet: Yann LeCun trained a convolutional neural network to read handwritten digits with >99% accuracy, proving CNNs work on real data.

2012 — AlexNet: Alex Krizhevsky's deep CNN won ImageNet by a 10.9% margin over the second place. It used GPUs for training — the moment deep learning became the default approach for computer vision.

From 2012, progress accelerated: AlphaGo beat the world Go champion in 2016, the Transformer architecture appeared in 2017, GPT-2 wrote coherent essays in 2019, and GPT-3 (175B parameters) demonstrated few-shot learning at scale in 2020. ChatGPT launched in November 2022 and reached 100 million users in two months.

What changed: data, compute, and the scaling hypothesis

Three ingredients drove the deep learning revolution:

Data: The internet created petabytes of labelled and unlabelled text, images, and video. ImageNet alone contained 1.2 million labelled images across 1,000 categories.

Compute: NVIDIA GPUs (designed for gaming) turned out to be ideal for the matrix multiplications neural networks require. Training time that took weeks on CPUs took hours on GPUs.

Algorithms: Backprop, dropout, batch normalisation, and the attention mechanism all made training deeper networks feasible.

The scaling hypothesis — that simply making models larger and training them on more data produces qualitatively better capabilities — has proven more durable than almost anyone expected. Today's frontier models have hundreds of billions of parameters and are trained on trillions of tokens.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to History of AI — Tutorial.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path