← Back to list
Lex Fridman PodcastPodcast15 Sep 2021Source: lexfridman.comHost: Lex Fridman

#221 – Douglas Lenat: Cyc and the Quest to Solve Common Sense Reasoning in AI

In plain words

This interview covers Douglas Lenat's 37-year Cyc project, which aims to give AI common sense by hand-coding millions of rules about how the world works. Lenat believes current AI like GPT-3 lacks real understanding and can make silly mistakes. He's optimistic about Cyc's approach, saying it's close to a tipping point where it can start learning on its own. Key mentions: Cyc (near completion, with tens of millions of common-sense rules); OpenCyc (a simplified version, often misunderstood); GPT-3 (risky—good 90% of the time, but unacceptable 1% errors in critical tasks).

AI SummaryAI-generated · may contain errors · verify against the original

Douglas Lenat discussed on the Lex Fridman podcast the core objective of his 37-year project, Cyc: addressing AI's common sense understanding problem through a commonsense knowledge base and reasoning mechanisms. The central argument is that despite progress in modern machine learning, building syst

~9 min full read · 8 sections
Deep Analysis

Here is the English translation of your investment research notes, following all specified rules.

At a Glance

Guest: Douglas Lenat, founder of the Cyc project, which aims to solve AI's common sense reasoning problem by manually constructing a commonsense knowledge base.

Main Theme: An exploration of the core challenges and methodology of the 37-year-old Cyc project, its relationship with current mainstream machine learning, and its potential value for future Artificial General Intelligence (AGI).

Core Thesis: Douglas Lenat argues that the key to building AGI lies in "priming the pump" of a "knowledge pump." This involves manually constructing a sufficiently large core knowledge base consisting of tens of millions of commonsense "assertions," enabling it to initiate a self-reinforcing learning loop that can automatically acquire and understand broader world knowledge.

Topic Sections

1. Common Sense: AI's "Brick Wall" and the Mission of Cyc

Lenat points out that in the 1980s, he and his colleagues, while developing various AI programs (e.g., natural language understanding, robotics, expert systems), all hit the same "brick wall": programs lacked common sense. They could perform specific tasks like a "smart dog," but "didn't truly understand what they were doing, saying, or being asked."

  • Definition of Understanding: Lenat likens "understanding" to solid, layered "ground" beneath one's feet. When encountering unexpected or novel situations (e.g., a garbage truck suddenly bursting open while driving), humans can rely on deep-seated common sense (e.g., "fast food trash bags usually contain Styrofoam and leftovers, you can drive over them") to make split-second decisions. AI lacks this deep support.
  • Core Task: Cyc's goal is to build this "ground" for computers—a commonsense knowledge base containing basic concepts and the rules of how the world works. This is not just about storing knowledge, but also about being able to efficiently retrieve and reason with it.
2. Knowledge Scale and Acquisition: Lessons from "Millions" to "Tens of Millions"

Lenat recalls a pivotal 1984 meeting with top scholars like Marvin Minsky and Alan Newell, where they collectively estimated the amount of knowledge required to cover common sense.

  • Initial Estimate: Using different methods (e.g., human long-term memory capacity, vocabulary size, representation of a single encyclopedia), everyone arrived at an estimate of roughly 1 million assertions/rules of thumb. This seemed achievable manually within 100 person-years.
  • Reality and Revision: After the project launched, they discovered the estimate was "off by an order of magnitude." The actual requirement was tens of millions of commonsense rules. Fortunately, due to the threat from Japan's "Fifth Generation Computer" project and the passage of the US National Cooperative Research Act, Cyc received far more funding and manpower (thousands of person-years) than anticipated, allowing it to continue to this day.
  • Knowledge Acquisition Methodology: To unearth this common sense that is never explicitly written down, the Cyc team employed a unique method: focusing on the "gaps" in text. They analyzed what authors assume readers already know, such as pronoun references and implicit logic between sentences (e.g., "Fred Smith robbed Third National Bank. He was sentenced to 20 years." — the steps of arrest, trial, and conviction are implicit). They also reverse-engineered common sense by analyzing "fake news" or contradictory statements (e.g., "Elvis is alive" vs. "Elvis's ghost").
3. Core Architecture: Local Consistency, Context, and "Left-Right Brain" Synergy

To handle the ubiquitous contradictions and exceptions in the real world, Cyc had to abandon global consistency in favor of a locally consistent architecture.

  • Context: The knowledge base is divided into independent contexts, akin to "plates." Each context is internally consistent, but different contexts may contradict each other (e.g., "the real world" vs. "the Looney Tunes cartoon world"). Context is a "first-class citizen" in Cyc, can be reasoned about like any other object, and forms a hierarchical structure (e.g., "Austin" is a sub-context of "Texas").
  • Balancing Efficiency and Expressiveness: Cyc splits the problem in two: the epistemological problem (what the system should know) and the heuristic problem (how the system reasons efficiently). It uses a clean, expressive higher-order logic language (the epistemological layer) and over 1,000 specialized, efficient "heuristic modules" (the heuristic layer). These modules, like a group of experts collaborating on a blackboard, solve problems in real-time while maintaining powerful expressiveness.
  • Synergy with Machine Learning: Lenat believes machine learning (e.g., deep learning) is analogous to the human "right brain," excelling at rapid pattern recognition, while the symbolic reasoning represented by Cyc is like the "left brain," excelling at slow thinking, causal reasoning, and explanation. The synergy of the two is key to achieving AGI. He cites a project with the NIH where Cyc constructed 10-20 step causal chains linking gene mutations found in Genome-Wide Association Studies (GWAS) to diseases, thereby filtering out noise and generating verifiable predictions.
4. From "Knowledge Pump" to AGI: Future and Challenges

Lenat firmly believes Cyc is close to completing the "priming" phase of the knowledge pump and is about to enter a phase of self-reinforcing automatic learning.

  • "Priming" Theory: Knowledge learning is a "surface area" problem. The more known knowledge, the larger the "surface" for learning, and the easier it is to learn new knowledge. Cyc's goal is to build a core large enough to learn automatically through reading, conversation, and experimentation.
  • Relationship with the Semantic Web and Open Knowledge: Lenat views the Semantic Web and knowledge graphs (e.g., Google Knowledge Graph) as "shallow" attempts. They can only represent simple binary relationships (e.g., "Fred's wife is Wilma"), but cannot handle complex logic like nested beliefs or intentions (e.g., "Juliet believes Romeo will think..."). Cyc's open-source version, OpenCyc, was originally intended to showcase the value of the full version, but many mistakenly believe OpenCyc represents the entirety of Cyc.
  • Commercialization and the Future: Cyc has shifted from primarily government funding to commercial applications, with clients including hospitals and energy companies. Lenat believes the future of AGI will not be "us vs. AI," but "us with AI." AI will act as a "teaching assistant" or "muse," helping humans think critically, identify blind spots, and accelerate scientific paradigm shifts.

Position Moves

Position Guest Sentiment Key Data
Cyc Bullish (Core project, nearing completion of "knowledge pump" priming) Contains tens of millions of commonsense assertions; over 1,000 heuristic reasoning modules; ~150 reasoning acceleration breakthroughs; 50-person full-time team working for decades
OpenCyc Neutral (A simplified projection of full Cyc, often misunderstood) Contains only simple binary relationships (knowledge graph) from Cyc, missing over 90% of the rules
GPT-3 Risk Warning (Lacks true understanding, makes commonsense errors) Performs well 90% of the time, but a 1% error rate is unacceptable for critical tasks (e.g., medical, driving)
IBM Watson Risk Warning (Makes absurd errors a human would never make) Impressive performance on Jeopardy!, but once answered "Ronald Reagan" for a 16th-century Italian politician
Mycin Bullish (Early expert system, provided explainable diagnoses) Developed by Ted Shortliffe, specialized in diagnosing blood infections (e.g., meningitis)

Judgments Worth Remembering

1. Common sense is AI's "brick wall," not its "ceiling." (Douglas Lenat) — Early AI systems could excel at specific tasks, but would fail completely when encountering unexpected situations requiring commonsense reasoning, due to a lack of basic world understanding.

2. Cyc's knowledge acquisition method is "reading the gaps," not "reading the text." (Douglas Lenat) — The team analyzes information authors assume readers already know but is not explicitly written (e.g., pronoun references, implicit logic between sentences); this is the core of common sense.

3. AGI requires the synergy of the "left brain" and "right brain." (Douglas Lenat) — Machine learning (right brain) excels at rapid pattern recognition, while symbolic reasoning systems like Cyc (left brain) excel at slow causal reasoning and explanation. Combining both is necessary for true intelligence.

4. The "Knowledge Pump" theory: The more you know, the faster you learn. (Douglas Lenat) — The ability to learn new knowledge is proportional to the "surface area" of the existing knowledge base. Cyc's goal is to build a core large enough to initiate a self-reinforcing automatic learning loop.

5. Global consistency is the enemy of building a large commonsense knowledge base. (Douglas Lenat) — The real world is full of contradictions and exceptions. Cyc solves this by dividing the knowledge base into multiple "locally consistent" contexts, allowing for inconsistencies between different contexts.

6. Cyc's "heuristic layer" is the key to its efficiency. (Douglas Lenat) — To achieve real-time reasoning on an expressive higher-order logic language, Cyc developed over 1,000 specialized heuristic modules that, like a group of experts on a blackboard, collaborate to quickly solve specific types of problems.

7. A good test for AGI is the depth of "why." (Douglas Lenat) — A truly intelligent system should be able to answer multiple rounds of "why" questions consecutively and provide arguments for and against a proposition, rather than just giving a single answer.

8. The ultimate value of AI is to help humans break free from the constraints of "paradigms." (Douglas Lenat) — Human scientists are often limited by the current scientific paradigm. A trustworthy AGI can help us step back, question assumptions taken for granted, and thus accelerate scientific discovery.