Douglas Lenat: Cyc and the Quest to Solve Common Sense Reasoning in AI | Lex Fridman Podcast #221 | Summary and Q&A

TL;DR

Psych, a project that aims to assemble common sense knowledge for artificial intelligence, is working towards creating a system that can think, reason, and understand the world based on general intelligence.

Key Insights

• 😊 Psych aims to solve the core problem of artificial intelligence by acquiring common sense knowledge and using it to understand the world. This will lead to advancements in AI systems and help us understand our own minds and be more rational and kind to each other.
• 😃 The goal of Psych is to assemble a knowledge base of common sense knowledge that spans the basic concepts and rules of how the world works. This is a challenging task but essential for achieving true AI.
• 😲 Psych addresses the limitations of existing AI programs, which lack common sense knowledge and true understanding of what they are doing. These programs can perform tasks, but they lack true comprehension.
• 🤔 Understanding is like a solid foundation that we rarely think about but rely on when faced with unexpected situations. Computers need this common sense knowledge to make split-second decisions in novel situations.
• 🔍 Philosophers have developed formal languages like predicate logic to enable mechanical procedures to derive logical conclusions from information. This helps computers understand and reason like humans.
• ️ The number of facts needed to capture the universe perfectly is an open question, but estimates suggest it could be on the order of millions or tens of millions. Psych has spent decades assembling these assertions to create a knowledge base.
• 🌐 Psych represents knowledge as a graph rather than a tree, allowing for more expressive representations and flexible linking between concepts. This enables effective reasoning and avoids the limitations of tree-like representations.
• 🤝 Psych's aim is to automate knowledge acquisition and expand the knowledge base using techniques like natural language understanding and knowledge editing tools. It aims to synergize machine learning with human insight to enhance the AI system.
• 🧩 The combination of machine learning and automated reasoning in Psych's knowledge base will provide a more comprehensive and robust AI. It will enable critical thinking and reasoning, allowing humans to augment their intelligence and solve complex problems.

Transcript

the following is a conversation with doug lennon creator of psych a system that for close to 40 years and still today has sought to solve the core problem of artificial intelligence the acquisition of common sense knowledge and the use of that knowledge to think to reason and to understand the world to support this podcast please check out our spon... Read More

Questions & Answers

Q: What is the core problem that Psych aims to solve in the field of artificial intelligence?

Psych seeks to address the challenge of acquiring common sense knowledge and using it to improve artificial intelligence's ability to think, reason, and understand the world.

Q: Why is the acquisition of common sense knowledge important for AI systems?

Common sense knowledge is crucial for AI systems to have a better understanding of the world, make accurate judgments, and perform tasks effectively. Without common sense, AI systems may lack true comprehension of their actions and responses.

Q: How does Psych's representation language help address the challenge of acquiring common sense knowledge?

Psych uses a formal language, such as predicate logic, to represent information in a way that enables a mechanical procedure to derive logical conclusions. This representation language helps AI systems reason and generate logical entailments like humans would.

Q: Can you explain the concept of understanding in the context of AI systems?

Psych describes understanding as a solid foundation of common sense knowledge that AI systems can rely on to accurately respond to queries, make reasoned decisions, and provide explanations. It involves leveraging knowledge, filtering out the useful information, and integrating it for effective decision-making.

Q: How does Psych plan to automate the expansion of its knowledge base?

Psych aims to integrate machine learning and natural language processing techniques to automate the acquisition of new knowledge. This includes using self-supervised learning methods and developing knowledge editing tools that enable users to expand the system's knowledge base easily.

Q: How does Psych approach handling exceptions and contradictions within its knowledge base?

Psych adopts a system of local consistency instead of global consistency to handle exceptions and contradictions. Contexts act as boundaries between different sets of knowledge, and small inconsistencies may exist at the context boundaries. The goal is to ensure each context is generally consistent, allowing for efficient decision-making while keeping the ability to reason about exceptions.

Q: Can machine learning techniques, like deep learning, be used to help expand Psych's knowledge base?

Yes, Psych aims to harness machine learning and natural language processing techniques to automate the expansion of its knowledge base. This involves leveraging machine learning models to extract relevant information from large datasets and incorporating it into the system's common sense knowledge.

Q: What is the core problem that Psych aims to solve in the field of artificial intelligence?

Psych seeks to address the challenge of acquiring common sense knowledge and using it to improve artificial intelligence's ability to think, reason, and understand the world.

Summary

In this conversation, Doug Lenat, creator of Psych, discusses the mission of his project to capture and incorporate common sense knowledge into artificial intelligence. He explains how common sense knowledge is crucial for AI systems to truly understand and reason about the world. Lenat shares how the process of capturing this knowledge has evolved and the challenges they faced in building a large knowledge base. He also talks about the importance of context and how Psych represents knowledge as a graph rather than a tree. Lenat believes that AI and humans working together can lead to greater intelligence and the ability to solve global problems.

Questions & Answers

Q: What is the goal of the Psych project?

The goal of the Psych project is to assemble a knowledge base that encompasses common sense knowledge about how the world works. This knowledge base would help AI systems to understand and reason about the world.

Q: What was the core problem faced by AI systems prior to Psych?

Prior to Psych, AI systems lacked common sense knowledge and general world understanding. They were limited to performing specific tasks or following predefined rules without comprehending the meaning or context behind their actions.

Q: How does anyone understand something?

Understanding can be thought of as a solid foundation of knowledge that underlies our thinking and reasoning. It is like the ground we stand on, where most of the time, we don't need to think about it. Understanding involves combining knowledge, inference, experience, and wisdom gained over time.

Q: How does common sense knowledge help in unexpected situations?

Common sense knowledge becomes crucial in unexpected situations where existing rules or knowledge might not be applicable. It provides a solid foundation for decision-making in novel or unforeseen circumstances. For example, when faced with a sudden obstacle while driving, common sense knowledge can help make split-second decisions.

Q: How do you leverage common sense knowledge to make split-second decisions?

Understanding involves the ability to leverage common sense knowledge and integrate it quickly for making split-second decisions. It's about effectively filtering out irrelevant information, selecting only useful parts, and combining them to reach a conclusive decision.

Q: How many pieces of common sense knowledge are needed for AI to have a comprehensive understanding?

The exact number of pieces of common sense knowledge needed to achieve comprehensive understanding remains unknown. Estimates from various experts have ranged from around one million to tens of millions. It is a complex question and may forever remain open-ended.

Q: How did the initial estimations of the number of pieces of knowledge change over time?

Initially, the estimation of the number of pieces of common sense knowledge required was believed to be around one to two million. However, after several years of research, it was realized that tens of millions of pieces of knowledge were needed.

Q: How did the Psych project approach acquiring knowledge efficiently?

Psych approached the acquisition of knowledge by analyzing the gap between sentences and the implicit assumptions made in texts. They also studied examples of fake news, humorous headlines, and contradictions to understand the kind of knowledge people assume and the reasons for disbelief or inconsistency.

Q: How was the knowledge base of Psych built over time?

The Psych project started in the 80s and initially focused on gathering a critical mass of general knowledge. By leveraging research funding and consortium efforts, they were able to assemble a large knowledge base comprising tens of millions of assertions or rules. This was an ongoing process that required thousands of person-years of effort.

Q: How does Psych ensure the generalizability and consistency of its knowledge base?

Psych constantly polices its knowledge base to ensure that assertions are as general as possible without losing accuracy. They use techniques like understanding assumptions, looking for inconsistencies, and developing generalized principles. Psych adopts a system of local consistency, where each context or domain can have its own consistent set of assertions, even if inconsistencies arise when contexts interact.

Q: Can contexts be formulated in a hierarchical structure?

Yes, contexts can be arranged hierarchically, similar to how general concepts can be specialized into more specific concepts. Each context can inherit knowledge from more general contexts, reducing the need to repeat information at every level.

Takeaways

Doug Lenat's Psych project aims to capture and incorporate common sense knowledge into AI systems. Common sense knowledge is essential for AI to understand and reason accurately about the world. The estimation of how much knowledge is needed lies somewhere between one to tens of millions of assertions. Psych's knowledge base is built by analyzing implicit assumptions, studying contradictions, and leveraging various examples. Psych represents knowledge as a graph, allowing contexts to interact while maintaining local consistency. The power of AI is the ability to combine human and AI intelligence, resulting in a mutually beneficial partnership to solve global challenges.

Summary & Key Takeaways

• Psych, created in 1984 and still active today, seeks to solve the core problem of artificial intelligence: acquiring common sense knowledge and using it to think, reason, and understand the world.

• Psych addresses the limitations of AI systems, which lack general world knowledge and understanding, by developing a formal language to represent information and using algorithms to generate logical conclusions.

• Understandability is a crucial aspect of AI intelligence, and psych aims to build a knowledge base that captures the common sense knowledge that humans possess.