A history of AI
128 milestones in the ideas, results, and people behind modern AI, from 1943 to today. Each cites its sources; 104 link to the lesson here that teaches the idea.
December 1943Foundations
McCulloch and Pitts model the neuron
Warren McCulloch and Walter Pitts published "A logical calculus of the ideas immanent in nervous activity" in the Bulletin of Mathematical Biophysics (volume 5, pages 115 to 133, December 1943). Starting from the all-or-none character of nervous activity, they treated neurons as threshold units whose firing can be described with propositional logic, showed that the behavior of every net can be described in such terms, and showed that suitable logical expressions can be realized by some net.
Why it mattered
Their binary threshold unit became the basic template for the artificial neuron used in later neural networks. The paper gave no procedure for learning connection strengths from examples; Hebb and Rosenblatt later took up learning.
- People
- Warren McCulloch, Walter Pitts
Sources
- PaperA logical calculus of the ideas immanent in nervous activity
- ArchiveWalter Pitts (IEEE Computer Society computer pioneers)
Learn the idea here
Foundations
1943 to 1955Mathematicians, neuroscientists, and engineers ask whether thinking can be described as computation, and propose the first models of neurons and of learning.
Symbolic AI and the first winter
1956 to 1979The field gets its name. Programs that manipulate symbols prove theorems and hold scripted conversations, and early neural networks learn simple patterns. Expectations outrun results, and funding contracts in the 1970s.
Expert systems and the second winter
1980 to 1992Rule-based expert systems reach commercial use, then interest fades late in the decade. Meanwhile backpropagation revives research on multilayer neural networks.
Statistical learning
1993 to 2011Methods with solid statistical footing, such as support vector machines, boosting, and random forests, dominate practice. Large datasets and graphics processors set up what comes next.
Deep learning
2012 to 2016Deep neural networks trained on GPUs win at image recognition, then spread to speech, translation, and games.
Transformers and scale
2017 to 2021The transformer and pretraining on large text collections produce language models whose loss falls predictably as they grow.
Generative AI
2022 to presentModels that generate text, images, and code reach the public. Tuning with human feedback, reasoning models, and open-weight releases reshape the field, and governments begin to regulate it.
How to read this timeline
The strip at the top is drawn to scale. Its contour lines show where the milestones listed here crowd together, the way a map shows a ridge. It measures this selection, not all AI research.
The track below gives every milestone room to be read, so its spacing is not to scale: years stretch where events are dense. Use the arrow keys to move through it, or Page Up and Page Down to change era.
Eras are named for the approach that dominated each period; the boundaries are conventions, and every era had work in other directions. Dates are as precise as the sources allow. Every entry cites a primary or official source (276 in all).
Looking for something to read next? See resources for textbooks, courses, and the papers behind these milestones.