AI Foundationspredict · compress · act
act V

The First Attempts

The first AI was not learning. It was logic, symbols, and hand-written rules — and it got remarkably far.

16

Rules All the Way Down

before this →Why Silicon Got Good at This
1956Dartmouth — the term "artificial intelligence" is coined
1957GPS — a general problem solver searches toward a goal
1965DENDRAL — first expert system, for chemistry
1972MYCIN — diagnoses blood infections better than some doctors
1980XCON — saves DEC tens of millions a year

From the 1950s to the 1980s, “AI” meant manipulating symbols. Represent the world as facts and rules, then reason over them. It is called symbolic AI, or GOFAI — “good old-fashioned AI.”

The machinery

Three ideas carried the era. Knowledge representation: encode facts as symbols (IF fever AND rash THEN ...). Search: explore a tree of possible states toward a goal — chess grandmasters fell to deep search. Expert systems: bottle a specialist’s rules and sell it. For narrow, rule-like domains, these worked, and sometimes made real money.

The brittleness

why rules alone hit a wall

Symbols are precise, and the world is not. A rule-base has no idea what to do with a typo, a blurred photograph, or a sentence it has never seen. Every exception you handle reveals three more. The knowledge has to be written by a human, by hand, forever.

the bottleneckExpert systems scale with the number of experts you can hire. Neural networks scale with the number of examples you can collect. That difference is the entire story of what happened next.
REPRESENTfacts and rules, written by hand
SEARCHexplore states toward a goal
BRITTLEone unhandled case → total failure

What it got right

Do not dismiss this era. Symbolic AI gave us search algorithms, theorem provers, databases, logic programming, and the idea that knowledge can be explicit and inspectable. Modern agent systems still use planning and explicit knowledge — they just pair it with a model that can handle the messy parts. The lesson is not “symbols were wrong.” It is “symbols alone are not enough.”

introduces →symbolic AIknowledge representationexpert systemsearch
← previousWhy Silicon Got Good at Thisnext →The First Winter