The founding era (1950–1956)
Artificial intelligence began as a question before it was a field. In 1950 Alan Turing asked whether machines could think and proposed his imitation game. Six years later, the 1956 Dartmouth workshop organised by John McCarthy, Marvin Minsky, Claude Shannon and others gave the field its name and its founding optimism. Early programs like the Logic Theorist and later ELIZA showed machines reasoning and conversing in narrow domains, and researchers confidently predicted human-level intelligence within a generation. That optimism defined — and later haunted — the decades that followed.
Symbolic AI and the first winter (1956–1980)
The dominant early approach was symbolic AI: encoding human knowledge as explicit rules and logic. It produced impressive demos but struggled with the messiness of the real world, common sense, and the sheer combinatorial explosion of possibilities. When progress stalled and influential funding reviews turned critical in the mid-1970s, money and enthusiasm dried up — the first AI winter. The lesson, repeated since, was that narrow lab successes do not automatically generalise to open-ended problems.
Expert systems and the second winter (1980–1993)
The 1980s brought a commercial revival through expert systems — rule-based programs that captured the knowledge of specialists in fields like medical diagnosis and equipment configuration. Companies invested heavily, but the systems were brittle, expensive to maintain, and unable to learn. When the specialised hardware market collapsed at the end of the decade, a second AI winter set in. Meanwhile, quietly, researchers were reviving neural networks — the connectionist approach — with the backpropagation algorithm, planting the seeds of what came next.
The deep learning breakthrough (1993–2017)
Through the 2000s, three trends converged: the internet generated enormous datasets, GPUs delivered cheap parallel computation, and learning algorithms steadily improved. The turning point was 2012, when a deep neural network called AlexNet crushed the ImageNet image-recognition benchmark, proving that “deep” multi-layer networks trained on big data could beat hand-crafted methods. A wave of successes followed in vision, speech, and games — including DeepMind’s AlphaGo defeating a world champion at Go in 2016. Deep learning had decisively replaced symbolic AI as the field’s centre of gravity.
The transformer and LLM era (2017–present)
In 2017 the paper Attention Is All You Need introduced the transformer, an architecture that uses attention to process whole sequences in parallel and scales extraordinarily well. It powered BERT (2018) and OpenAI’s GPT series. As models and training data grew, capabilities emerged that no one had explicitly programmed. The public inflection point came in late 2022 with ChatGPT, which brought conversational AI to hundreds of millions of people almost overnight. The years since have been defined by ever-larger and more capable models — GPT-4, Claude, Gemini, and open-source families like Llama — plus a return to the field’s oldest debates about reasoning, safety, and what intelligence really is. After seventy years of booms and winters, AI is now woven into everyday software.