1.4 Learning apprentices and continuous knowledge acquisition

The learning-apprentice tradition makes the feedback loop operational. LEAP, developed for VLSI design, was conceived as an interactive knowledge-based assistant that acquired rules by observing the ordinary problem-solving steps of expert users. When a user overrode its advice and manually refined a circuit, the system could treat that intervention as evidence of a rule it should have possessed [ Mitchell et al. , 1985 ] .

The later account of apprentice-based knowledge acquisition separated at least two learning obligations. LEAP learned feasible problem reductions and also learned control knowledge for choosing among competing reductions. Its explanation-based procedures operated with incomplete domain theories rather than presuming that every useful dependency had already been encoded [ Mahadevan et al. , 1993 ] . This distinction anticipates the present separation between acquiring arrows that make an operation possible and acquiring observers or preferences that determine when the operation should be used.

The historical architecture also contains a lesson for current frontier models. Knowledge acquisition need not be isolated in a separate training phase. A useful assistant can learn while participating in the task, provided that it records the teacher’s contributions, reconstructs their explanatory role, and preserves uncertainty about what has not yet been established. Language models greatly expand the available interface for explanations and corrections, but fluent dialogue does not by itself produce a coherent task theory. The acquired claims must still be typed, related, and tested.