History of AI: from Turing to language models

In short: AI has been a research field since the 1950s with booms and setbacks (“AI winters”). Today’s breakthrough rests on three things: lots of data, graphics cards and the transformer architecture.

In more detail: The ideas are old, but only enough data and compute made them practically usable.

In detail

• 1950: Alan Turing asks “Can machines think?” (Turing test).
• 1956: Dartmouth conference, the term “artificial intelligence”.
• 1958: Perceptron by Frank Rosenblatt, a single learning neuron.
• 1970s/80s: rule-based expert systems, then the first AI winter (expectations too high).
• 1986: backpropagation becomes known (Rumelhart, Hinton, Williams).
• 1997: Deep Blue beats world chess champion Kasparov. LSTM is introduced.
• 2012: AlexNet wins ImageNet with GPUs, start of the deep learning boom.
• 2013: Word2Vec. 2014: GANs, seq2seq with attention.
• 2016: AlphaGo beats Lee Sedol.
• 2017: “Attention Is All You Need”, the transformer architecture.
• 2018–2020: BERT, GPT-2, GPT-3 (175 bn parameters), scaling laws.
• 2022: diffusion models (Stable Diffusion, DALL·E 2), ChatGPT makes AI mass-market.
• 2023–2025: open models (Llama), multimodality, reasoning models, agents, EU AI Act.
• 2024: Nobel prizes for Hinton/Hopfield (physics) and Hassabis/Jumper (chemistry, AlphaFold).

See also: How AI works: overview and map, Transformer architecture