The Origin of Artificial Intelligence: A Grounded History

AI writing tends to come in one of two registers: breathless hype or performative skepticism. Neither is particularly useful. What actually happened is more interesting than both: a seventy-year run of mathematicians, engineers, and a few obsessive hobbyists trying to make machines do things that looked like thinking, using whatever tools were on hand at the time. This document follows that thread, including the people, the dead ends, and the two separate occasions the field nearly gave up on itself.

The Precursor: Theoretical Foundations (1930s-1940s)

The Question of Machine Intelligence

In 1936, Alan Turing published On Computable Numbers, describing a theoretical device, later called the universal computing machine, capable of carrying out any computation that could be precisely specified. He wasn't trying to build a thinking machine. He was trying to answer a narrower question: what, exactly, can be computed at all? That question turned out to be the foundation everything else in this document sits on.

Turing's wartime work on code-breaking, including his contributions to the Bombe machine, showed that this kind of formal computation wasn't just theoretical. Machines really could carry out tasks like deduction and pattern-matching at scale that had until then been assumed to require a human mind.

Cybernetics and Feedback Systems

Norbert Wiener spent the 1940s developing cybernetics, the study of how systems, biological or mechanical, regulate themselves through feedback. A thermostat is a crude example: it senses a condition and corrects for it. Wiener's insight was that this same loop, generalized, might explain something about how animals and eventually machines adapt to their environment. It gave the field a vocabulary for adaptive behavior before anyone had built a system sophisticated enough to need one.

The Birth of AI as a Discipline (1950-1956)

The Turing Test

In 1950, Turing published Computing Machinery and Intelligence, proposing what's now called the Turing Test: if a human evaluator, conversing blindly with both a machine and another human, can't reliably tell which is which, the machine passes. Turing was careful not to claim this proved consciousness or understanding. It only gave researchers something concrete to test against, instead of arguing philosophy indefinitely.

The Dartmouth Conference (1956)

The phrase "Artificial Intelligence" didn't exist until the summer of 1956, when John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon organized a workshop at Dartmouth College built on a genuinely bold premise: that any aspect of intelligence could, in principle, be described precisely enough for a machine to simulate it. Nobody at Dartmouth had proof of this. It was a bet, and the field spent the next several decades finding out how much of that bet would pay off.

Early AI: Symbolic Reasoning and the First Programs (1950s-1960s)

Logic Theorist (1956)

Allen Newell and Herbert Simon built the Logic Theorist, generally credited as the first AI program, to prove mathematical theorems using symbolic logic rather than brute computation. It wasn't just solving math. It was demonstrating that deductive reasoning itself could be mechanized.

ELIZA (1966)

Joseph Weizenbaum's ELIZA simulated a psychotherapist through little more than pattern matching and clever rephrasing of the user's own words back at them. ELIZA understood nothing. What unsettled Weizenbaum was how many people didn't seem to care, an early, uncomfortable preview of how easily people project understanding onto a system that has none.

The Limits of Symbolic AI

Programs like ELIZA and Logic Theorist worked because their problems were narrow and well-defined. Real-world ambiguity broke them. By the late 1960s, it was clear that rule-based systems alone weren't going to scale to anything resembling general intelligence, and funding for AI research collapsed. It was the first of two periods now known as AI Winters.

The Rise of Expert Systems and Machine Learning (1970s-1980s)

Expert Systems

The 1970s response was to go narrower and deeper: encode a human expert's specific knowledge into an explicit set of rules. Two examples defined the era:

• MYCIN, built at Stanford across the decade, diagnosed bacterial infections and recommended treatment by working through a rule base.
• XCON, deployed by Digital Equipment Corporation in the 1980s, configured computer orders automatically. It was one of the first AI systems to save a company real money.

Both worked well inside their domain and fell apart the moment reality drifted outside it. That brittleness, not stupidity but an inability to generalize, is the defining flaw of the whole expert-systems era.

The Emergence of Machine Learning

Arthur Samuel started building a self-improving checkers program at IBM in 1952, and by 1959 had published the paper that coined the term machine learning. The distinction from symbolic AI mattered: instead of hand-coding rules, Samuel's program got better by playing itself and learning from the outcome. It was a quiet, unglamorous idea at the time. It's the idea that eventually ate the entire field.

The Second AI Winter

By the mid-1980s, expert systems had become a real industry, complete with specialized Lisp machine hardware built to run them. Then the market collapsed. General-purpose workstations got cheap and capable enough to do the same job, and the expert systems themselves turned out to be expensive to maintain: every rule needed a human to update it as the world changed. Funding agencies, already once burned, pulled back hard. This second winter ran into the early 1990s and taught the field the same lesson twice: hand-coded knowledge doesn't scale.

The Data Revolution and the Revival of Neural Networks (1990s-2000s)

The Internet and Big Data

The thing that finally broke the pattern wasn't a new algorithm. It was the internet, which produced more raw data than any hand-built rule system could ever hope to encode manually, and gave data-driven approaches something to actually learn from.

IBM's Deep Blue (1997)

Deep Blue's 1997 defeat of world chess champion Garry Kasparov is often misremembered as a triumph of "thinking." It wasn't. Deep Blue combined brute-force search with hand-tuned evaluation heuristics to check millions of positions per move, closer to industrial-scale calculation than to insight. It proved machines could beat humans in structured, rule-bound domains. That's a narrower claim than it sounds, but it was a real one.

The Revival of Neural Networks

Networks of simple units predate the field by over a decade. In 1943, Warren McCulloch and Walter Pitts published a paper describing neurons as logical switches: on or off, firing or not. Wire enough of them together and the resulting network could, in theory, compute anything formal logic could express. Nobody built one. It was a mathematical proof, not a machine.

Frank Rosenblatt changed that in 1958 with the perceptron, a physical device that learned to classify patterns by adjusting its own connection weights. The press coverage was enormous and mostly wrong. In 1969, Marvin Minsky and Seymour Papert published Perceptrons, proving the architecture couldn't handle certain basic problems. They were right about the math, and neural network research went quiet for nearly two decades.

The 1986 popularization of backpropagation, by Rumelhart, Hinton, and Williams, showed how to train networks with multiple layers, solving the core limitation Minsky and Papert had identified. But the hardware to make that method count wouldn't exist for another two decades. The idea was ready long before the infrastructure was.

The Modern Era: Deep Learning and Specialization (2010s-Present)

The Breakthroughs of the 2010s

Three things converged in the 2010s: far more data, GPUs that made large-scale training feasible, and a genuinely new architecture. The first two were resource problems working themselves out. The third, the transformer, was a real idea.

Key Milestones

• AlexNet (2012) dramatically outperformed prior approaches to image recognition and effectively kicked off the deep-learning era.
• AlphaGo (2016), from DeepMind, beat a world champion at Go, a game with too many possible positions for brute-force search to handle the way Deep Blue had handled chess. AlphaGo learned through deep reinforcement learning instead.
• The Transformer (2017), introduced in Google's "Attention Is All You Need," replaced older sequence-processing methods with a mechanism that weighs relationships across an entire input simultaneously. This is the architecture underneath essentially every large language model in use today. It's arguably the single most consequential idea in AI's second half-century.

The Scaling Era (2018-2022)

What followed the transformer wasn't a new breakthrough so much as a blunt empirical observation: these models kept getting better as researchers threw more data and more parameters at them, with no obvious wall in sight. In 2018, Google's BERT showed that a single model pre-trained on a large text corpus could be fine-tuned for dozens of different tasks, from answering questions to classifying sentiment. It was the practical proof that one general model could replace many specialized ones.

By 2020, researchers had quantified the pattern into scaling laws: model performance improves predictably as you increase compute, data, and parameter count. That finding turned AI from a research discipline into a capital expenditure race. Training costs went from thousands of dollars to hundreds of millions. That observation produced foundation models, systems trained once on broad data, then adapted to specific tasks through prompting rather than being rebuilt from scratch each time. It marked a real departure from the AlphaGo-and-AlexNet model of AI, where one system did one narrow thing well.

The Conversational Turn (2022 onward)

Foundation models existed for years before most people noticed. What changed in 2022 wasn't really the underlying technology. It was reinforcement learning from human feedback (RLHF), which steered raw model output toward being useful and instruction-following rather than just statistically plausible, wrapped in an interface anyone could type into. That combination, not a research breakthrough, is what took AI from a research topic to a mainstream one.

Where Things Stand Now

Today's systems, from Anthropic, OpenAI, Google, and others, are still, mechanically, extensions of the 2017 transformer: trained to predict patterns, refined through human feedback. What's changed is scale, task range, and a growing ability to use tools and act across multi-step problems, sometimes called agentic AI. Whether that adds up to genuine generality or just very fluent pattern-matching is a live argument, and this document isn't going to pretend to settle it. What holds up is the throughline: every stage of this history solved a narrower problem than "intelligence" writ large, using whatever mathematical and engineering tools existed at the time. That's still true.

The Core Idea: AI as a Tool

AI automates tasks that once required human perception, reasoning, or judgment. It is not, and has never been, at any point in this history, about manufacturing consciousness. The story is one of pragmatism: solving whatever problem was in front of the field with whatever tool actually worked, whether that was formal logic, statistics, or a neural network.

Conclusion

Nothing here was sudden. AI's history is a long grind, interrupted twice by winters that forced the field to abandon approaches that didn't scale, and pushed forward by people solving concrete problems rather than chasing a grand vision. The Logic Theorist, AlphaGo, and today's transformer-based systems are testaments to human ingenuity, not to any will or awareness on the machine's part. Worth remembering, especially now, that the field's actual roots are reason, data, and the practical application of ideas, not the mythology that's grown up around it since.

If You Want to Go Deeper

This is a high-level history, and it skips over a lot to stay that way. A handful of the original papers behind it are worth reading directly if any of this stuck with you.

Turing's two papers are the obvious starting point: On Computable Numbers (1936) for the theoretical foundation, and Computing Machinery and Intelligence (1950) for the Turing Test itself, still more nuanced than the pop-science version of it. The Dartmouth workshop proposal from 1956 is short and worth skimming just to see how confident the founders of the field were about a problem they'd barely started to define. Arthur Samuel's 1959 paper on his checkers program is the origin of the term "machine learning," and it's a good reminder of how unglamorous the idea looked at the time. Rumelhart, Hinton, and Williams's 1986 paper on backpropagation is the one that made modern neural networks trainable, decades before the hardware caught up. And Attention Is All You Need (2017), the paper that introduced the transformer architecture, is the single most important thing to read if you want to understand why today's AI looks the way it does.

None of these require a technical background to get something out of, though a couple get dense in the middle.