Cyc: history's forgotten AI project

🌈 Abstract

The article discusses the history and current status of the Cyc project, one of the most ambitious artificial intelligence projects in history. Cyc is a four-decade-long effort to codify the common-sense knowledge that is the foundation of human reasoning, with the goal of enabling human-like reasoning in machines.

🙋 Q&A

[01] Precursors

1. What were the key issues in AI research that led to the creation of the Cyc project?

• The main issue was how to program machines with common sense, which was seen as a key challenge in achieving genuine intelligence in AI systems.
• Previous AI programs like AM and EURISKO had exhibited flashes of intelligence, but were limited by their hard-coded heuristics and inability to learn and adapt on their own.
• Researchers believed that a machine with a large knowledge base of common-sense facts and rules could overcome these limitations and form the foundation for a new generation of more effective AI systems.

2. What was the proposed scale and approach for building this common-sense knowledge base?

• Researchers calculated that a machine would need to know several million rules before it could begin to learn on its own.
• Assembling such a knowledge base was estimated to require 2,000 person-years of effort, which was seen as feasible with the right resources.
• The approach was to systematically codify basic facts and rules about the world that humans take for granted but are not explicitly written down.

[02] Cyc is born

1. How was the Cyc project initiated and funded?

• In 1984, the Microelectronics and Computer Technology Corporation (MCC), a research consortium of 10 US companies, launched the Cyc project as its flagship initiative.
• MCC provided Cyc with a half-billion dollar budget and hundreds of employees, allowing project leader Doug Lenat to pursue his vision on a large scale.

2. What was the structure and approach of the Cyc knowledge base?

• Cyc's knowledge base is a collection of "frames" (or units) with slots for properties and entries for their values, organized into a global ontology.
• On top of the frames, Cyc has a constraint language for expressing logical concepts and an inference engine for making deductions and answering queries.
• The knowledge was originally entered by hand by "ontological engineers" who identified common-sense facts and rules from analyzing natural language text.

[03] Cyc's development and challenges

1. How did Cyc's development progress in its early years?

• By 1989, Cyc's knowledge base had reached 1 million "pieces of data", including 50,000 individual units and 6,000 collections.
• The project encountered and dealt with various "representational thorns" - difficulties in representing real-world knowledge within the system.
• The guiding philosophy was pragmatism over elegance, with practical solutions to age-old philosophical problems.

2. What were some of Cyc's applications and partnerships in the later years?

• Cyc was used by the Cleveland Clinic to assist medical researchers in answering ad hoc questions, reducing the time from a month to less than an hour.
• Cyc also partnered with the US intelligence community to build a "terrorism knowledge base" that could synthesize information from different sources.
• Cycorp, the company that spun off from MCC, released subsets of the knowledge base to researchers, but the core product remained proprietary.

[04] Cyc in the age of LLMs

1. How did the rise of machine learning and deep learning impact the perception of Cyc?

• As the field of AI shifted towards neural networks and deep learning, which proved successful at previously intractable problems, Cyc's rule-based, symbolic approach increasingly looked like an anachronism.
• Cyc was seen by many in the AI community as a cautionary tale of tremendous effort wasted on a misguided approach, overshadowed by the breakthroughs of deep learning.

2. What is the current status and future potential of Cyc?

• Despite the criticism, Cyc has survived for 40 years and continues to be funded through commercial contracts, employing 50 technical staff.
• While Cyc's impact has not been revolutionary, the article suggests that it could potentially complement large language models (LLMs) like ChatGPT, which are fluent but often inconsistent and inaccurate, by providing a chain of reasoning that can be audited.
• The article concludes that the course of progress in AI is unpredictable, and that rule-based systems like Cyc may have a role to play again in the future.