Paradigm Shifts
aihistoryproject.org
Paradigm Shifts
Eight decades of AI research, grouped into the paradigms that defined them — symbolic AI rising and collapsing, the statistical turn, and the deep-learning era that swallowed the field.
VIEW:
Each decade normalised to 100% — paradigm mix, free of the uneven item counts per decade.
Foundations & Cyberneticspeak 1940s
The theoretical groundwork — computability, feedback, and the mind as a machine.
Symbolic AI & Knowledgepeak 1960s
Intelligence as explicit rules and symbols — logic, expert systems, knowledge bases.
Statistical ML & Datapeak 2000s
The statistical turn — learning patterns from data rather than hand-coded rules.
Neural Nets & Deep Learningpeak 2010s
Connectionism and its revivals — perceptrons, backpropagation, deep architectures.
Generative AI & LLMspeak 2020s
Scale, pretraining and generation — language models, agents, and reasoning.
Perception, Language & Roboticspeak 2010s
The applied frontier every paradigm attacked — seeing, speaking, and moving.
Compute & Infrastructurepeak 1940s
The substrate — the machines, chips, and systems each era could run on.
Safety, Ethics & Governancepeak 2020s
The consequences thread — alignment, bias, policy, and who answers for it.
What the Shape Shows
30% → ~0
Symbolic AI’s arc
Takes nearly a third of the 1970s, then all but vanishes. One 2010s entry inherits it — knowledge graphs — and it survives as infrastructure rather than as a research programme.
1990s
The statistical turn
Statistical ML & Data reaches its high point as learning from data displaces hand-coded rules.
Twice
Neural nets return
A 1980s backpropagation revival, a dip through the 2000s, then the 2010s deep-learning surge.
40%
The generative era
Generative AI & LLMs go from nothing before the 1980s to the largest share of the 2020s.
Derived from the tags on all 451 timeline items · 426 carry at least one paradigm tagHover a band or a card to isolate it