SHRDLU understood blocks, not open conversation

The conversation

Terry Winograd built SHRDLU at the MIT Artificial Intelligence Laboratory between 1968 and 1970, as Winograd's own project page states, describing a program that "carried on a simple dialog (via teletype) with a user, about a small world of objects (the BLOCKS world) shown on an early display screen." The same page preserves the demonstration exchange used to showcase it: a person types "PICK UP A BIG RED BLOCK," the program answers "OK" and executes the move; asked to "GRASP THE PYRAMID," it replies "I DON'T UNDERSTAND WHICH PYRAMID YOU MEAN" when more than one pyramid is in the scene, and it asks again when a pronoun such as "it" is ambiguous.

What the documents show

Winograd's dissertation, issued as MIT AI Technical Report 235 in February 1971, is described in its own abstract as a system that "answers questions, executes commands, and accepts information in normal English dialog," built around "a detailed model of the knowledge needed by a simple robot having only a hand and an eye." The report frames the program as combining syntactic parsing, a model of the block-world scene, and deductive reasoning about the objects and the robot arm's own actions and plans, so a command like finding a block taller than the one currently held requires SHRDLU to track what it is doing as well as what it is being asked to do next.

The system boundary

The boundary in both documents is the world itself: SHRDLU's grammar, inference, and memory apply only to blocks, boxes, and pyramids on a simulated tabletop, plus the hypothetical hand that moves them. Winograd's abstract describes the program "requesting clarification when context and physical knowledge prove insufficient for comprehension" — a defined fallback, but one that operates entirely inside the block world's own vocabulary. Neither document claims the same grammar-plus-inference approach would scale to unrestricted conversation; the project page instead links to the source code and later ports precisely because the domain is small enough to reimplement.

Where it fails

SHRDLU's demonstrations were widely retold as evidence a computer could hold a normal conversation, a description neither source supports. The technical report is explicit that the competence belongs to the closed toy domain, not to open-ended English. A builder should distinguish a system that reasons well inside a narrow, fully specified world from one that reasons well about the world generally.

• Is the system's fluency being demonstrated inside a domain small enough that every object and rule is already known to the program?
• What does the system do when a request falls outside its modeled world, and is that different from a clarifying question inside it?
• Would the same architecture, unmodified, handle a domain the original demonstration never covered?

More than fifty years on, the gap between a closed simulated world and general dialogue is still the first thing to check in any claim built on SHRDLU's example.