SHRDLU

What it was

SHRDLU was an early natural-language understanding program developed by Terry Winograd at MIT between 1968 and 1970; Wikipedia's infobox lists a 1967 release date for it.^[1] It ran on a DEC PDP-6 computer with a DEC graphics terminal, and was written in Micro Planner and Lisp.^[1]

The setting was deliberately small: a simplified blocks world, essentially a virtual box filled with blocks. Inside it, a user could carry on a conversation with the computer — moving objects, naming collections, and querying the state of the world.^[1]

What it could do

The world was described with a vocabulary of roughly 50 words. Among them were nouns such as block and cone, verbs such as place on and move to, and adjectives such as big and blue.^[1] Small as that vocabulary was, it was enough to support a genuine back-and-forth: the user gave instructions, asked questions, and the program replied in English.

The world also contained basic physics, so that blocks would fall over; this simulation ran independently of the language parser.^[1] SHRDLU kept a basic memory to supply context, so that after a sentence about a green cone, the shorter phrase "take the cone off" could still be understood.^[1]

It could answer questions about what was possible and impossible in its world, and questions about history and causes; one explanation ended with the line "BECAUSE YOU ASKED ME TO."^[1] It could also remember names given to objects or arrangements: it accepted a definition of a "steeple" and then built one.^[1]

How the knowledge was represented

The work behind SHRDLU was written up as MIT AI Technical Report 235, "Procedures as a Representation for Data in a Computer Program for Understanding Natural Language", dated 24 August 1970 in Wikipedia's further reading; MIT's DSpace record lists the same report as AITR-235, issued January 1, 1971.^[1][2] The report itself is a revised version of a dissertation submitted to the Department of Mathematics on August 24, 1970, supervised by Seymour A. Papert.^[3]

The title states the central design choice: many kinds of knowledge were represented in the form of procedures rather than as tables of rules or lists of patterns.^[3] In other words, knowing what a word meant was bound up with what the program could do with it. The report's abstract describes a system that answers questions, executes commands and accepts information in ordinary English dialogue, using semantic information and context to understand discourse and to disambiguate sentences; the program is given a detailed model of the knowledge a simple robot with only a hand and an eye would need.^[2]

Reception and later view

The demonstrations made a strong impression, and according to Wikipedia the program led other AI researchers to excessive optimism — optimism that was soon lost when later systems ran into the ambiguity and complexity of more realistic settings.^[1]

Winograd himself later became one of the sharpest critics of that impression. In a 1991 interview he said he had carefully worked through the demonstration line by line, and that there was no attempt to make the program usable by other people; the pressure, as he described it, was for something you could demo.^[1] He used the phrase "Potemkin villages" for demonstrations that look good but lack enough structure to work more generally.^[1]

That combination — a compact, carefully bounded world that produced strikingly fluent behaviour, and a later recognition of how much of that fluency depended on the boundary — is why SHRDLU is still cited. It showed what a program could do with language when the domain is small and fully modelled, and it set the terms for the argument over how far such demonstrations carry.