Bet #1: Write Down the Rules (Symbolic AI)

In January 1956, Herbert Simon walked into his class at the Carnegie Institute of Technology in Pittsburgh with news. In the version most histories repeat, he told his students: “Over Christmas, Allen Newell and I invented a thinking machine.”

The machine was a program called the Logic Theorist, written by Newell, Simon and a RAND programmer named Cliff Shaw. Its job was to prove theorems in logic, the kind found in Principia Mathematica, the giant book in which Bertrand Russell and Alfred North Whitehead tried to build mathematics from pure logic.

It worked. The Logic Theorist proved 38 of the first 52 theorems in chapter 2 of the book. For one of them, theorem 2.85, it found a proof shorter and more direct than the one Russell and Whitehead had published.

Proving theorems was something only very clever people did. Now a machine was doing it by following rules. If logic could be mechanised in a few months, how far away could the rest of thinking be?

That feeling launched the first big bet on general intelligence. Intelligence is the manipulation of symbols by rules, so if we write down the right symbols and the right rules, we get a mind. For about thirty years, it was the main way people tried to build AI. This is the story of that bet, and of how much of it still runs inside the AI you use today.

The idea: thinking is symbol shuffling

A symbol is something that stands for something else. The word “cat”, the letter x in algebra, a chess piece on a diagram. A rule says how symbols can be combined or changed: if x + 2 = 5, then x = 3. If all men are mortal and Socrates is a man, then Socrates is mortal.

The symbolic bet says this is what thinking is. Your mind holds symbols and rules for reasoning about them. A computer can do the same, so given the right symbols and rules, it can think.

Newell and Simon later stated it as a formal hypothesis in their 1975 Turing Award lecture, published in 1976:

“A physical symbol system has the necessary and sufficient means for general intelligent action.”

“Necessary” means you can’t be generally intelligent without symbol manipulation. “Sufficient” means symbol manipulation is all you need. That is a bet on AGI in one sentence.

The philosopher John Haugeland later gave it a nickname: GOFAI, “Good Old-Fashioned AI”: logic, rules, search and hand-built knowledge, as opposed to the learning-based AI of the next part.

Dartmouth, 1956: the bet is placed

AI got its name in a funding proposal.

On 31 August 1955, four researchers, John McCarthy (Dartmouth College), Marvin Minsky (Harvard), Nathaniel Rochester (IBM) and Claude Shannon (Bell Labs), wrote to the Rockefeller Foundation. They proposed “a 2 month, 10 man study of artificial intelligence” at Dartmouth College in the summer of 1956. It’s generally cited as the first use of the phrase “artificial intelligence.”

The proposal’s central sentence is the clearest statement of the symbolic bet ever written:

“The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.”

Every aspect. So precisely described. If you can describe it exactly, you can program it.

It was also optimistic about the schedule: “a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer.”

The summer didn’t produce that advance, but it produced a field. Newell and Simon brought the Logic Theorist, by most accounts the only working AI program anyone showed that summer.

The golden years: proofs, puzzles and a chatbot

The late 1950s and 1960s were heady. Programs did things nobody had seen machines do before.

General Problem Solver (1957 onwards)

Newell, Shaw and Simon followed the Logic Theorist with a more ambitious program: the General Problem Solver, or GPS. The name tells you the ambition. Instead of one kind of problem, it would solve any problem you could state formally.

Its key trick was means-ends analysis. Compare where you are with where you want to be. Find the biggest difference. Pick an action that reduces it. If you can’t do that action yet, make “being able to do it” a new sub-goal, and repeat.

It’s a real insight, still taught today. GPS solved logic puzzles and textbook problems like the Towers of Hanoi. But “general” turned out to mean “general across small, neat, fully described puzzles.”

The predictions

The early successes produced some famous forecasts:

• In 1958, Simon and Newell predicted that within ten years a computer would be the world chess champion.
• Simon wrote in 1960, and repeated in his 1965 book The Shape of Automation for Men and Management, that “machines will be capable, within twenty years, of doing any work a man can do.”
• Minsky wrote in his 1967 book Computation: Finite and Infinite Machines that within a generation “the problems of creating ‘artificial intelligence’ will be substantially solved.” (A popular reworded version of this line circulates; Minsky later said he’d been misquoted, so this is the wording from the book.)

A computer did become world chess champion, in 1997, 29 years late.

ELIZA (1966)

At MIT, Joseph Weizenbaum wrote ELIZA, a program that held typed conversations. Its most famous script, DOCTOR, imitated a certain kind of therapist who mostly reflects your words back. Type “I’m worried about my mother” and it might answer “Tell me more about your family.”

ELIZA had no understanding at all. It matched keywords against patterns and filled in templates. But people confided in it anyway. Weizenbaum reported that his own secretary, who knew exactly what it was, asked him to leave the room so she could talk to it in private.

That shook him. He spent much of his career warning how easily people credit machines with understanding, most famously in his 1976 book Computer Power and Human Reason. In a 2025 Turing test study, ELIZA, used as a baseline, still fooled judges 23% of the time.

SHRDLU (1968 to 1970)

At MIT, Terry Winograd built SHRDLU, the most impressive demo of the era. A user typed English commands to a simulated robot arm in a small world of coloured blocks, pyramids and boxes. You could say “put the green cone on the red block” and then “take the cone off”, and it would understand that “the cone” meant the one you’d just mentioned. It could answer questions about what it had done and why.

It felt like real language understanding. But its world and vocabulary were tiny. These small artificial worlds were called microworlds, and the hope was to master small worlds and then join them up into the big one.

That hope is where the bet first hit a wall.

Wall #1: the real world is too big

The problem has a name: combinatorial explosion. Symbolic programs worked by search: trying sequences of rules or moves and checking which ones lead to the goal. With a few blocks, the possibilities are manageable. Add more objects, actions and steps, and every step multiplies the options, so the number of paths grows exponentially. Methods that solved toy problems in seconds became hopeless on realistic ones.

The second problem was knowledge. SHRDLU knew everything about its world because its world was tiny. A program that works in a kitchen, a hospital or a conversation about politics needs to know an enormous amount of ordinary stuff that nobody ever writes down.

The Lighthill report (1973)

In 1972, the British Science Research Council asked the applied mathematician Sir James Lighthill to review AI research. His report, “Artificial Intelligence: A General Survey”, came out in 1973 and was brutal. His most quoted line: “In no part of the field have the discoveries made so far produced the major impact that was then promised.”

Lighthill argued that AI had worked on simplified problems and failed to face the combinatorial explosion that real problems bring. The BBC then televised a debate at the Royal Institution in London: Lighthill against Donald Michie, John McCarthy and Richard Gregory.

The report became the basis for the British government ending support for AI research at most UK universities. US funders also grew more demanding about practical results. The mid-1970s became what’s now called the first AI winter: less money, fewer jobs, and “AI” became a slightly embarrassing word.

The comeback: expert systems

The bet came back in a narrower, more commercial form. Researchers like Edward Feigenbaum at Stanford drew a lesson from the winter: general reasoning tricks weren’t the secret. Knowledge was. An expert is an expert because they know a lot about one field. So capture that knowledge as rules, and you get an expert in a box.

These were called expert systems. Each had two parts: a knowledge base, hundreds or thousands of IF-THEN rules written with human experts, and an inference engine that chained the rules together to reach a conclusion.

MYCIN (1970s)

At Stanford, Edward Shortliffe and colleagues built MYCIN to recommend antibiotics for serious infections. Its rules took the form IF these lab findings and patient facts, THEN suggestive evidence of this organism, and because medicine is uncertain, it attached certainty factors to conclusions instead of yes or no.

In a 1979 study published in JAMA, eight infectious disease experts blindly rated treatment plans for ten meningitis cases. MYCIN’s choices were rated acceptable 65% of the time. The plans of five Stanford faculty specialists scored between 42.5% and 62.5%.

On paper, it was as good as the doctors. It was never used to treat patients. Fitting it into real hospital work, keeping it current and handling the questions of a computer advising doctors was a harder problem than getting the answer right.

XCON (1980)

The big commercial success came from a less glamorous job. Digital Equipment Corporation (DEC) sold VAX computers in many configurations, and getting each parts list right was tedious and error-prone. John McDermott at Carnegie Mellon built R1, later called XCON, to do it. It went into use at DEC in 1980.

By 1986, it had processed about 80,000 orders with reported accuracy of 95 to 98%. The most cited estimate is that it saved DEC around $25 million a year, though figures vary between sources. Its rule base grew from hundreds of rules to thousands.

XCON proved that rules could make money. In the early 1980s, corporations raced to build expert systems, a small industry of AI companies and specialised hardware sprang up, and AI was hot again.

Japan’s Fifth Generation (1982)

Governments joined in. In 1982, Japan’s Ministry of International Trade and Industry (MITI) launched the Fifth Generation Computer Systems project, run by a new institute called ICOT. The plan was to build massively parallel computers designed for logic programming, using the language Prolog, as the foundation for knowledge-based AI. The budget was in the region of ¥50 to 57 billion over about a decade, a few hundred million US dollars at the time, depending on the source and what you count.

It alarmed the US and Europe into launching their own programmes. It produced real research in parallel logic programming, but not the intelligent machines or the market it promised, and it wound down in the early 1990s.

Wall #2: brittleness

The boom ran into three problems that still apply to any rule-based system you build today.

1. Brittleness. An expert system knew its narrow area and nothing around it. Ask it something just outside its rules and it didn’t degrade gracefully, it failed, often confidently. It had no common sense to fall back on.
2. The knowledge acquisition bottleneck. Every rule had to be pulled out of a human expert’s head through long interviews. Experts often can’t explain what they know. They disagree with each other. And the world changes, so the rules have to be updated forever.
3. Maintenance. As rule bases grew into thousands of rules, they interacted in ways nobody could predict. Adding one rule could quietly break another.

At the same time, the specialised “Lisp machines” built for AI work were overtaken by cheaper ordinary workstations from companies like Sun, and that market collapsed around 1987. Funding dried up again. The late 1980s and early 1990s became the second AI winter.

Expert systems didn’t vanish. They became ordinary software: rule engines for credit checks, tax calculations and fraud detection. They just stopped being called AI, which is a pattern you’ll see again.

The biggest single bet: Cyc

Doug Lenat, a Stanford-trained AI researcher, argued that expert systems failed because they lacked common sense: the millions of ordinary facts every child knows. Water is wet. If you let go of something, it falls. No expert system knew any of this, so they made absurd mistakes the moment they stepped outside their rules.

His answer: write it all down. In 1984, at a research consortium in Austin, Texas called MCC, he started Cyc (from “encyclopedia”). A team would hand-encode common sense in formal logic, fact by fact. Lenat hoped that once the base was big enough, Cyc could learn the rest itself by reading. In 1994, the project spun out into a company, Cycorp.

Cyc is the purest test of the symbolic bet there is, because it’s the one project that didn’t stop. In a paper Lenat co-wrote with Gary Marcus and posted in July 2023, they described the result: a four-decade, roughly 2,000 person-year effort that produced a knowledge base of tens of millions of hand-written assertions and rules, with a reasoning engine that gives a complete, auditable chain of logic for every answer.

That last part is real and valuable: today’s neural networks can’t show their reasoning that way. But the takeoff never came. Cyc didn’t learn the rest by reading, and it didn’t become a general intelligence. Lenat died on 31 August 2023, weeks after that paper appeared. His final argument, with Marcus, was that trustworthy AI would need to combine large language models with an explicit knowledge-and-reasoning system like Cyc.

As of October 2026, Cycorp is still in business. Its website presents Cyc as “machine reasoning AI” built on codified common sense and markets it for healthcare operations work like hospital billing and staffing forecasts. Critics have written obituaries for the AGI ambition; supporters note that a small company kept a 40-year project alive on real contracts. Both are true: Cyc became useful software, not a mind.

Deep Blue: the bet’s greatest victory, and what it really was

On 11 May 1997, in New York, IBM’s Deep Blue won the sixth and final game against world chess champion Garry Kasparov and took the match 3.5 to 2.5. It was the first time a computer had beaten a reigning world champion in a match under standard tournament conditions. Simon’s 1958 prediction had come true, decades late.

It’s remembered as a triumph of AI. It’s more accurately a triumph of search.

Deep Blue ran on an IBM supercomputer with custom chess chips and could evaluate about 200 million positions per second. It looked many moves ahead and scored the resulting positions with an evaluation function: a big hand-designed formula, split into about 8,000 features of a position, whose weights were tuned partly by analysing thousands of master games. Grandmasters helped prepare its opening book.

It didn’t learn chess the way AlphaZero later would, by playing itself. Its knowledge was largely put there by people, its strength came from speed, and it could do nothing except play chess. So Deep Blue proved both halves of the symbolic story at once. Search plus hand-built knowledge can be superhuman inside a closed world with exact rules. And none of it carried over to anything else. In DeepMind’s grid from Part 1, Deep Blue is a narrow system at Level 4: extraordinary depth, zero breadth.

The deepest wall: common sense and the frame problem

In 1969, John McCarthy and Patrick Hayes described the frame problem. When you describe an action in logic, you have to say what changes. But you also somehow have to say what doesn’t change. If a robot paints a box red, its position doesn’t change, the other boxes don’t change, and so on, forever. Writing out every non-effect is impossible, and assuming nothing else changes is sometimes wrong. Knowing what’s relevant without checking everything turned out to be fiendishly hard.

Philosophers took the problem wider. The critic Hubert Dreyfus had argued since the 1960s, in his 1972 book What Computers Can’t Do, that human expertise isn’t mostly rule-following at all: it’s intuition and context built from being a body in a world, and you can’t write it down as rules. AI researchers were often furious with him; many of his points aged well.

The roboticist Hans Moravec noticed a related irony, now called Moravec’s paradox: it was relatively easy to get computers to do “hard” adult things like logic and chess, and very hard to give them the “easy” skills of a toddler, like seeing, walking and picking things up. What we find hard, we do consciously, step by step, so it’s easy to write as rules. What we find easy, we do without knowing how.

That is the core reason the symbolic bet didn’t reach general intelligence. Most of what we know, we can’t say. And a strategy built on saying it hit that limit.

The legacy: symbolic AI never went away

This isn’t a story of pure failure. The symbolic bet lost the race to general intelligence but won many smaller ones, and its ideas run in systems you use every day.

• Search and planning. The SRI robot Shakey, built from 1966 to 1972, gave us the A* search algorithm and the STRIPS planner. Descendants of these ideas route your delivery driver, plan logistics, schedule factories and move characters in video games.
• Knowledge graphs. In May 2012, Google launched its Knowledge Graph with the slogan “things, not strings”: a database of more than 500 million objects and 3.5 billion facts and relationships about them. The fact box you see when you search a famous person is symbolic AI.
• Rules and logic in business. Tax software, insurance underwriting, compliance checks and fraud rules are expert systems in all but name. Logic solvers verify chip designs and check critical software for bugs.
• Tool use by today’s AI. Large language models are bad at exact arithmetic, precise lookups and guaranteed-correct logic, the very things symbolic systems are good at. So modern AI agents call symbolic tools: a calculator, a code interpreter, a database query, a search index, a planner. The neural network handles the fuzzy language; the old-fashioned symbolic tool gives the exact answer. I wrote about how that works in What an “AI Agent” Actually Is.
• Hybrids. In January 2024, DeepMind published AlphaGeometry in Nature: a language model that suggests ideas, paired with a symbolic deduction engine that checks them. It solved 25 of 30 olympiad geometry problems in its test set. That kind of marriage is the subject of Part 7 of this series.

The symbolic camp also left the right questions: how to represent knowledge, how to make reasoning explainable, how to know what’s relevant. Neural systems are still working on all three.

Where the bet stands in October 2026

Almost nobody now believes you can reach general intelligence by only writing down rules by hand. The common-sense problem and the knowledge bottleneck were too big, and learning from data (the next bet) solved perception and language in ways rules never did.

But “symbolic AI is dead” is wrong too. Symbolic components are everywhere in 2026 AI systems as tools, verifiers, planners and knowledge bases. And a real camp, including Gary Marcus and many neurosymbolic researchers, argues that general intelligence will need explicit symbols and reasoning combined with learning. The debate is no longer “rules or learning?” It’s “how much structure, and where?”

Scoreboard

Key ideas

• The symbolic bet: thinking is symbol manipulation, so precisely described knowledge plus rules should produce intelligence. The 1955 Dartmouth proposal said so in one sentence.
• Early wins were real but small: Logic Theorist proofs, GPS puzzles, ELIZA chats and SHRDLU’s blocks world all worked inside tiny, tidy worlds.
• Two winters followed two booms: combinatorial explosion and the Lighthill report in the 1970s, then brittle and costly expert systems in the late 1980s.
• Cyc was the bet’s purest test: about 2,000 person-years and tens of millions of hand-written assertions produced useful, explainable software, not a general mind.
• Deep Blue was search, not learning: superhuman in one closed world, with nothing that carried over.
• The deepest wall was common sense: the frame problem and Moravec’s paradox show that most human knowledge is never written down.
• The legacy is everywhere: planners, knowledge graphs, rule engines and the exact tools that today’s AI agents call.

Sources: J. McCarthy, M. L. Minsky, N. Rochester and C. E. Shannon, “A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence” (31 August 1955); Carnegie Mellon, “AI trailblazers” (CMU magazine) and Wikipedia, “Logic Theorist” (citing Pamela McCorduck, Machines Who Think) for Simon’s January 1956 remark and the 38-of-52 result; A. Newell and H. A. Simon, “Computer Science as Empirical Inquiry: Symbols and Search” (Communications of the ACM, 1976); H. A. Simon and A. Newell, “Heuristic Problem Solving: The Next Advance in Operations Research” (Operations Research, 1958); Quote Investigator on Simon’s “any work a man can do” (1960, reprinted 1965) and Minsky’s “within a generation” (1967); Wikipedia, “ELIZA” and “SHRDLU”; Terry Winograd, Understanding Natural Language (1972); Cameron R. Jones and Benjamin K. Bergen, “Large Language Models Pass the Turing Test” (arXiv 2503.23674, 2025); Sir James Lighthill, “Artificial Intelligence: A General Survey” (Science Research Council, 1973) and Wikipedia, “Lighthill report”; University of Edinburgh AIAI, BBC “Controversy” debate archive (1973); Yu et al., “Antimicrobial selection by a computer: a blinded evaluation by infectious diseases experts” (JAMA, 1979) and Buchanan and Shortliffe, Rule-Based Expert Systems (1984); Wikipedia, “Xcon”; Wikipedia, “Fifth Generation Computer Systems” and the ICOT final evaluation archive (Waseda University); Doug Lenat and Gary Marcus, “Getting from Generative AI to Trustworthy AI: What LLMs might learn from Cyc” (arXiv 2308.04445, 2023); Gary Marcus, “Doug Lenat, 1950-2023” (Substack, 2023); Cycorp website (accessed October 2026); Michael Tsai, “Obituary for Cyc” (April 2025), summarising Yuxi Liu’s essay on Cyc; IBM, “Deep Blue” (IBM history) and Wikipedia, “Deep Blue (chess computer)”; J. McCarthy and P. J. Hayes, “Some Philosophical Problems from the Standpoint of Artificial Intelligence” (Machine Intelligence 4, 1969); Stanford Encyclopedia of Philosophy, “The Frame Problem”; Hubert Dreyfus, What Computers Can’t Do (1972); Hans Moravec, Mind Children (1988); Google, “Introducing the Knowledge Graph: things, not strings” (16 May 2012); Trinh et al., “Solving olympiad geometry without human demonstrations” (Nature, January 2024).

Related: previous, What Does “AGI” Even Mean?; next, Bet #2: Let It Learn (Neural Networks); and What an “AI Agent” Actually Is.