Knowledge and Reasoning

Knowledge and reasoning is the pairing at the center of classical, symbolic artificial intelligence: first write down what an agent knows in a formal language, then let a general inference procedure draw conclusions from it. The two halves are deliberately separate. Representation decides what can be said, and reasoning decides what follows from what has been said, so that a new fact can change behaviour without any change to the program. This design philosophy, sometimes called the knowledge-based approach, dominated the field from the late 1950s into the 1980s and still shapes knowledge graphs, ontologies and planning systems.

John McCarthy set out the program in his 1958 paper "Programs with Common Sense," presented at a symposium at the National Physical Laboratory in Teddington and published in 1959, which described a hypothetical system called the Advice Taker. The Advice Taker would be given its knowledge as declarative sentences in a formal language and would deduce what to do from them, so that it could be improved by being told things rather than by being reprogrammed. McCarthy's example involved a person working out how to get to the airport by chaining together simple facts about walking, driving and being somewhere. The proposal separated knowledge from control, anticipated logic programming, and made common sense, not specialist expertise, the central problem.

The difficulty that grew out of this ambition was the common sense problem: ordinary life draws on a vast body of unstated knowledge about objects, people, space, time and cause, and nobody had a systematic way to write it down. Expert systems of the 1970s, such as DENDRAL and MYCIN, sidestepped it by confining themselves to narrow domains where rules could be collected from specialists. In 1974 Marvin Minsky proposed frames, structured packages of expectations about stereotyped situations, and Allen Newell's 1982 paper "The Knowledge Level" argued that an agent could be described by what it knows and what goals it has, independent of how that knowledge is encoded. Related work on semantic networks, scripts, and later description logics explored the trade-off between expressiveness and tractable inference.

The frame problem names a specific obstacle: how to state compactly what does not change when an action occurs. McCarthy and Patrick Hayes introduced it in their 1969 paper "Some Philosophical Problems from the Standpoint of Artificial Intelligence," which also presented the situation calculus. If a robot moves a block, a naive logical description needs separate axioms saying that the colour of the block, the position of the other blocks and the walls of the room are all unchanged, and the number of such axioms explodes. The problem pushed researchers toward non-monotonic logics, including McCarthy's circumscription and Reiter's default logic, both from 1980, in which conclusions can be withdrawn when new facts arrive. Philosophers later generalized the term to the wider problem of deciding what is relevant, which is related but not identical.

Cyc, begun in 1984 by Douglas Lenat at the Microelectronics and Computer Technology Corporation in Austin, was the boldest attempt to solve the common sense problem by brute-force encoding. The idea was that a team of knowledge engineers could enter the everyday facts and rules that underlie understanding, in a formal language with a large inference engine, until the system could learn the rest from reading. The project continued for decades, later under the company Cycorp, accumulating millions of assertions and a rich ontology that was used in commercial and government applications. It never became the general intelligence its founder hoped, and it is a standard example of the knowledge acquisition bottleneck, yet it also forced detailed engineering of questions that others only discussed.

Large language models invert the classical arrangement: knowledge is absorbed implicitly into parameters from text, and reasoning happens as generation, not as logical inference over explicit facts. The strengths are breadth and robustness to messy input; the weaknesses are fabrication, difficulty in guaranteeing consistency, and opacity about what the system knows. Current research tries to recombine the halves, through retrieval from external stores, knowledge graphs such as the one Google announced in 2012, tool use, and neuro-symbolic hybrids in which a neural model proposes and a symbolic checker verifies. A common misconception is that symbolic knowledge bases failed entirely; they remain dominant in places where facts must be auditable. Another is that the frame problem is only a robotics curiosity, when in fact any system must decide which consequences to track.

Read knowledge and reasoning as AI's founding division of labour, store what is known and derive what follows: it delivered auditable systems and a sharp statement of the common sense problem, and its unfinished business, supplying the world's knowledge at scale, is what both learned models and hybrid systems now try to meet.