Textbook of AI

SHRDLU, developed by Terry Winograd at MIT in 1968–1970, was the most impressive natural-language understanding system of its decade. It conversed with a user about a "blocks world" of coloured blocks and pyramids on a virtual table, parsing complex sentences ("Find a block which is taller than the one you are holding and put it into the box"), executing the implied actions in its simulated world, and answering follow-up questions including those requiring knowledge of its own past actions ("Why did you do that?"). The name SHRDLU is the second column of a Linotype keyboard, an old printers' joke for "garbled".

SHRDLU integrated three then-novel components: a procedural-grammar parser that encoded grammatical knowledge as procedures (in PLANNER, an early logic-programming language), a model of the blocks world that supported physical reasoning, and a discourse component that maintained a context of recent referents. The combination produced dialogue that, in the bounded domain of its blocks world, exceeded the linguistic competence of any prior system.

The system became, retrospectively, an emblem of the micro-worlds approach to AI, the idea that competence in narrow toy domains would scale to competence in the real world. The approach did not scale; SHRDLU's limited domain hid an enormous amount of hand-crafted knowledge that was specific to it. Winograd himself moved away from AI in the 1980s, partly in response to the realisation that micro-worlds techniques would not generalise.