# What is the Cyc project, and can you really teach a computer common sense by hand?

Date: 2020-02-23 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-the-cyc-project/

**Answer:** Cyc is a project, started by Douglas Lenat in 1984, to write down the everyday common sense that people never bother to say out loud, in a logical language a computer can reason with. After more than 35 years and over 1,000 person-years of hand-entered knowledge, it is the largest attempt ever made to give a machine common sense by teaching it fact by fact.

If you tell a friend, "I put the milk back in the fridge," you don't need to add that the milk is still in its carton, that the fridge is in the kitchen, or that the milk will not float up to the ceiling. Your friend knows all that. Computers don't. Cyc is the most ambitious attempt to fix this by brute force: to write down, in a form a computer can reason with, the millions of small truths that people take for granted.

The project began in July 1984 at the Microelectronics and Computer Technology Corporation (MCC) in Austin, Texas, a research consortium formed by about two dozen large American companies. It was led by Douglas Lenat, a computer scientist from Stanford University. Since January 1995 it has been run by Cycorp, a company in Austin that Lenat heads. The name comes from "encyclopedia" and is pronounced like "psych."

## Why Lenat started it

By the early 1980s, AI had a pattern that frustrated many researchers. A program would work impressively on a narrow problem and then fall apart on anything slightly unexpected. Expert systems, programs that captured a specialist's rules for jobs such as medical diagnosis, were the commercial success of the day, and they showed the problem clearly.

In a 1985 article in *AI Magazine*, Lenat and his colleagues Mary Prakash and Mary Shepherd named two bottlenecks. The first was **brittleness**: software breaks when it meets a situation its builders never imagined. The second was the sheer human labor of feeding knowledge into each new system. They argued that machine learning, automatic programming and understanding natural language had all struggled to scale up for the same reason: none of them had a large base of real-world knowledge to stand on.

So Lenat proposed building that base once, for everyone. At a 1983 meeting at Stanford, he, Alan Kay and other leading researchers estimated the job would take between 1,000 and 3,000 person-years of work. That was far too big for a university lab. It became possible when MCC, formed partly as an American response to Japan's ambitious Fifth Generation computer project, made Cyc its flagship effort.

## How Cyc works

Cyc has three main parts.

- **A language, CycL**, for stating facts and rules in strict logic. It looks somewhat like the Lisp programming language and grew far more expressive over its first five years, because ordinary sentences such as "Fred does not own a dog" or "Every American has a mother" turned out to need more than simple lists of facts.
- **A knowledge base** of concepts and statements, entered mostly by hand by trained "ontological engineers." By 1994 it held about 100,000 concepts and about a million assertions.
- **An inference engine**, the program that combines stored facts to reach new conclusions. Instead of one general reasoner, Cyc uses many specialized modules, each good at a particular kind of problem. Cycorp's own site now describes more than 1,100 of them working together.

Two design choices matter a lot. Most knowledge in Cyc is true only *by default*: parents usually love their children, but there are exceptions. And the knowledge is divided into "microtheories," separate contexts that each have to be consistent on their own. That way a statement can hold inside one context, such as a work of fiction or a single field of knowledge, without being treated as true everywhere.

## A worked example: how Cyc answers a question

Here is how Cyc can answer a question no one typed in directly. Suppose its knowledge base contains these two entries:

| Step | What Cyc knows or does | Type |
|---|---|---|
| 1 | Every tree is a plant | Stored fact (a category link) |
| 2 | Plants die eventually | Stored rule |
| 3 | Question: Do trees die? | New query |
| 4 | Tree is a kind of plant, and the rule about plants applies to every kind of plant | Inference |
| 5 | Answer: Yes, with the two steps shown as its reasoning | Explained conclusion |

The same mechanism handles richer cases. Given a photo caption, "Someone watching his daughter take her first step," Cyc can infer that the picture probably shows a smiling adult. It chains together defaults: parents love their children; a first step is a big accomplishment; a loved one's accomplishment makes you happy; happy people smile; only adults have children. And it can list those steps, so a person can check exactly why it reached its answer.

> **Try it:** Pick an ordinary sentence, such as "She left her umbrella at the restaurant." Write down every unstated fact a reader needs to understand it (what an umbrella is for, that restaurants are places you visit, that she probably wants it back). Most people stop at ten; Cyc tries to have them all.

## What came of it

Cyc kept growing. By 2017 its knowledge base held about 1.5 million concepts and about 24.5 million hand-written assertions, which took more than 1,000 person-years to write. Cycorp released a free, partial version called OpenCyc in 2002 and a fuller ResearchCyc for academics in 2006; OpenCyc was withdrawn by 2017, and ResearchCyc stopped being supported in December 2019. Cycorp now sells Cyc to businesses and government agencies, saying it represents over thirty years of engineering and hundreds of millions of dollars in research.

Opinions about Cyc split sharply. Marvin Minsky praised Lenat for being one of very few researchers to take common sense seriously. The machine-learning scientist Pedro Domingos has called the project a "catastrophic failure," arguing that no one can type in enough knowledge by hand and that Cyc cannot grow on its own. Both views agree on one thing: common sense is the hard part.

## How does Cyc's approach compare with language models like GPT-2 and BERT?

They sit at opposite ends of AI. Cyc is **symbolic**: people write explicit facts and rules, and the computer reasons with them step by step. GPT-2, which OpenAI released in full in November 2019, and Google's BERT are **statistical**: neural networks that read enormous amounts of text and learn patterns of which words tend to follow or surround others. Nobody types a single fact into them.

Each side has what the other lacks. A language model can write fluent paragraphs on almost any topic, but it cannot show you why it said something, and it sometimes states plain nonsense with confidence. Cyc can explain every conclusion, but it knows only what someone has entered, and it cannot read a sentence and simply pick up a new idea.

Many researchers in early 2020 suspect the future lies in combining the two: systems that learn from data at scale but can also draw on structured knowledge and explain their reasoning. Cyc remains the largest structured store of common sense anyone has built, and the clearest test of how far hand-built knowledge can go.

## Sources

- CYC: Using Common Sense Knowledge to Overcome Brittleness and Knowledge Acquisition Bottlenecks: https://doi.org/10.1609/aimag.v6i4.510
- Cyc: https://web.archive.org/web/20200124210050/https://en.wikipedia.org/wiki/Cyc
- Cyc (Cycorp home page): https://web.archive.org/web/20200223154514/https://www.cyc.com/