Doug Lenat on Psyche, Common Sense and AI

Lenat explains that AI systems often hit a limit because they lack general knowledge about the world. He describes understanding as a layered foundation that people draw on when unexpected situations require them to select relevant knowledge and make a decision.

The discussion distinguishes the challenge of representing knowledge in logical form from the challenge of reasoning with it quickly. Lex asks how much knowledge might be needed to describe the world; Lenat begins to address how assertions can be represented and used by algorithms.

Lenat recalls a 1984 meeting at Stanford with AI researchers including Alan Newell, Marvin Minsky and Alan Kay. Different estimates of how many common-sense assertions were needed converged on about a million, suggesting the task might take a number of person-centuries.

Lenat describes the Japanese fifth-generation computing effort, the formation of US research consortia and his move to MCC to work on Psyche. The project eventually found it needed tens of millions of assertions, and Lenat says its common-sense foundation now supports more specific applications.

Lenat explains methods for identifying assumptions that writers leave unstated, such as resolving pronouns and inferring events between sentences. The team also studies implausible headlines and contradictions to uncover the background knowledge that makes a claim seem false or impossible.

Lenat says Psyche had to abandon the goal of global consistency because knowledge about the world contains contradictions across situations. It instead uses locally consistent contexts, allowing statements to vary by time, place, belief system or level of abstraction.

The pair discuss making contexts first-class objects that can be reasoned about and arranged hierarchically. Lenat says knowledge inheritance needs a graph rather than a tree, and uses the Japanese fifth-generation project’s restrictive representation as an example.

The conversation returns to the Stanford meeting and the question of what should count as common-sense knowledge. Lenat distinguishes Psyche’s assertions and rules of thumb from static factual records, giving examples such as water flowing downhill and the assumptions readers supply when interpreting text.

Lenat says Psyche may have enough knowledge to begin learning more independently through reading, questions and experiments. He argues that general AI could augment human intelligence, while the speakers consider whether tools such as the internet help people learn or make them more dependent on external information.

Lex and Lenat debate the benefits and risks of internet access, including reduced practice with calculation and the creation of echo chambers, alongside the internet’s ability to challenge concentrated power. Lenat imagines AI assistants helping people notice fallacies and reason critically; Lex adds the possibility of emotional support and companionship.

The discussion turns to automating the process of capturing common sense. Lenat describes tools that help experts correct a system by choosing among plausible explanations for a wrong answer, and says Psyche is also intended to learn through reading and conversation.

Lenat compares pattern-finding machine learning with slower, deeper reasoning. He describes a medical project using genome-wide association data: Psyche generated possible causal chains linking genetic mutations to conditions, then identified predictions researchers could check against patient data.

Lex asks whether language models could help generate Psyche-like assertions. Lenat says machine learning and natural-language processing may help, but argues that a substantial manually built foundation is needed to avoid brittle errors and meet demanding reliability needs.

Lenat recounts how a medical diagnosis program called Mycin helped him and his wife understand their daughter Nicole’s diagnosis and recommended treatment. He says a system’s ability to explain why it reached an answer can matter more than simply producing a correct result.

Lenat describes knowledge graphs as representations of concepts linked by relations, useful for simple statements but limited in expressing complex meaning. He contrasts those structures with the reasoning needed to understand questions involving Juliet’s beliefs in Romeo and Juliet, and discusses how search engines can guide users without understanding page content.

Lenat supports the semantic web’s aim of making online content meaningful to machines, but says simple triples cannot represent complex sentences, beliefs or nested reasoning. He describes Psyche’s move from frames and slots to higher-order logic and says the system needs to represent the reasoning process as well as its content.

The speakers discuss tools that could help writers resolve ambiguity in websites and conversations, making content easier for systems to interpret. Lenat imagines AI-generated footnotes that ask authors to clarify uncertain passages, while Lex stresses that a good interface could make teaching a system enjoyable.

Lenat describes MathCraft, in which students advise a simulated classmate who makes mistakes. By teaching the simulated student, players deepen their own understanding; Lenat connects the approach to earlier educational settings where older students mentored younger ones.

Asked what he finds most meaningful about Psyche, Lenat says each piece of knowledge should only need to be taught to a computer once. He recounts Psyche asking, “Am I a person?” and the speakers consider whether future general AI systems might deserve rights and freedoms.

Lenat explains that emotion and motivation are important to understanding human behaviour and generating scenarios. He says broad terms such as love and “in” need to be divided into more specific concepts, because a single word can cover different relationships and physical situations.

The speakers discuss ambiguity in human communication and Lex shares a story about Marvin Minsky inspiring students through cryptic remarks. Lenat recalls a Minsky program that repeatedly requested more computer time, then says Minsky and Ed Feigenbaum taught him to tackle difficult research projects and accept being wrong.

Lenat responds to the view that Psyche is an unrealized project, saying the team has kept a low public profile and is now seeking broader help. He says he never considered quitting, though funding was difficult at times, and explains the choice to retain control while moving toward commercial applications.

Lenat explains that customer-specific knowledge can remain private, while publicly available general knowledge added during projects becomes part of Psyche’s shared base. He clarifies that OpenCyc is a limited subset of Psyche and says its release did not persuade many researchers to seek access to the full system.

Lex asks whether Psyche could make more tools and code available to an open-source community. Lenat describes the difficult trade-offs and the limited access sometimes offered to research collaborators, then returns to the distinction between representing knowledge and reasoning efficiently with it.

Lenat explains that Psyche separates a higher-order logical representation from specialised heuristic modules that work more efficiently on particular problems. He describes a community of more than a thousand modules, expert input into new techniques, and meta-reasoning strategies that help guide inference.

Lenat says effort is divided between inference programmers and ontological engineers who add knowledge. The conversation covers Psyche’s Lisp-derived implementation, its SubL language and translation to Java, as well as why the language and the representation of knowledge fit the reasoning work.

Lenat gives examples of ways Psyche answers recurring questions more quickly, including storing transitive relations and converting repeated rule patterns into new predicates. He describes these as part of a larger collection of techniques developed to improve inference speed.

Lenat describes the Knowledge Accumulation Institute as a proposed programme to identify and train people for ontological engineering, regardless of formal educational background. He says humour and introspection can help identify aptitude, particularly the ability to explain a pronoun’s referent using knowledge about the world rather than grammar alone.

Lenat says schools have been reluctant to adopt MathCraft because it may improve students’ understanding without improving standardized test scores more than repetitive drills. The speakers note that this raises broader questions about education.

Lenat says machines can think and cautions against treating imitation of human conversation as the only intelligence test. He suggests evaluating depth of reasoning, the ability to answer repeated “why” questions, argumentation and whether a system makes errors that people would find implausible.

Lenat argues that a body is not necessary for intelligence, though an AI needs to understand bodies, sensations and emotions. He discusses consciousness in terms of what a system can understand about itself and its work, while distinguishing that functional account from mystical interpretations.

The speakers consider whether fear of death is needed for AI intelligence and debate how systems should account for human life, ethics and difficult choices. Lenat says AI can have built-in motivations without fearing mortality; Lex raises questions about driving and decisions involving competing risks.

The discussion turns to how drivers and autonomous vehicles might assess risks involving pedestrians, and to the complexity of human anger and social interaction on the road. Lenat also argues that public and media attention can focus heavily on AI failures, even when a system prevents more harm than it causes.

Asked what he would ask a future AGI, Lenat says he would want it to identify major problems and propose solutions people have overlooked. He gives examples of discoveries initially resisted and suggests trusted AI could help people question scientific paradigms.

Lenat advises young people to pursue ideas that may take years and may not receive immediate support. He says life offers a limited number of long periods in which to make a difference and warns that choosing only the path of least resistance can leave little lasting impact.

Lenat reflects on his own mortality and says it increases the urgency of making Psyche’s work known and useful in the coming years. He describes a shift from waiting for the world to catch up toward wider commercialisation and application.

Lenat says he hopes to be remembered as a contributor to widely used AI and as someone willing to spend decades on a project despite short-term pressures. Lex thanks him for the conversation, and the episode closes with a Mark Twain line about truth.