Cyc: Douglas Lenat's thirty-year project to hand-write common sense
Sometime in the early 1980s, a medical diagnostic program returned a confident diagnosis of measles for Douglas Lenat’s Chevrolet. The program had learned patterns of disease presentation well enough to pass tests. It had not learned that cars lack circulatory systems. It was working exactly as designed.
In July 1984, Lenat set out to fix this. His base of operations was the Microelectronics and Computer Technology Corporation — MCC — in Austin, Texas, a research consortium pooling American industrial effort against Japan’s fifth-generation computing initiative, which had Washington worried about losing the AI race. Expert systems, the dominant technology of the moment, were brilliant within narrow bounds and brittle the instant they crossed them. The fix, Lenat reasoned, was not better inference engines. It was a foundation — the ten million unstated facts that literate adults share and never have to articulate. He would write them all down.
The project was called Cyc, from encyclopedia, and its premise was easy to state. If you told a computer that Napoleon died in 1821 and that Wellington was greatly saddened, a system with common sense would understand that death is permanent, that people grieve each other, and that the two men had a shared history. Without that background, the sentences were just tokens. Lenat and a team of philosophers, linguists, and computer scientists began encoding knowledge in a custom language called CycL, an extension of formal logic. Back-of-envelope calculations with Marvin Minsky, Allen Newell, and John McCarthy put the total effort at 1,000 to 3,000 person-years.
The execution was harder than the arithmetic. Lenat estimated eighty minutes per usable rule, accounting for eliciting the knowledge from a domain expert, encoding it in CycL, and debugging the cascade of inferences it triggered. The debugging was the work. An early attempt to encode “gravity carries X downward” produced the inference that gravity, lacking friends, could not rescue anyone from drowning — because another rule specified that a rescuer must have friends. Every rule was a thread. Pulling one pulled everything.
By 1994, Cyc held 100,000 terms and a million axioms. By the time Lenat died in August 2023, the knowledge base had grown to more than 1.5 million terms and 25 million assertions, at a cost estimated at $200 million and 2,000 person-years. The CIA, Goldman Sachs, and the Cleveland Clinic each used pieces of it. Whether it ever worked at the scale Lenat imagined is still disputed: Pedro Domingos called it a “catastrophic failure”; Lenat called it a foundation the industry chose to walk around rather than stand on.
He was probably right about both. What Cyc made undeniable was that the gap between machine intelligence and human knowledge is not a deficit of data or processing power — it is a structural absence, the ten-million-fact substrate of common sense that every adult carries without noticing. The deep-learning wave of the 2010s chose to learn that substrate statistically, from massive corpora, rather than write it down by hand. Whether it arrives at the same places Lenat was heading is the question the field is still working out.
Sources
- Cyc — Wikipedia — founding at MCC in July 1984, CycL language, axiom and term counts through 2017, Cycorp spin-off, Domingos criticism.
- He Taught AI the Facts of Life — Stanford Magazine — Lenat’s measles-diagnosis anecdote, project approach and collaborators.
- Cyc — Yuxi Liu — eighty minutes per rule, gravity-and-friends inference bug, Napoleon/Wellington test case, 1,000–3,000 person-year estimates with Minsky, Newell, and McCarthy.