SHRDLU

In the AI Story

The blocks world contained perhaps two dozen objects. Every object had a shape (block, pyramid, box), a color (red, blue, green, yellow, white, orange), and a size (big, small). The spatial relationships were limited: on, in, behind, to the left of, to the right of. Every word had exactly one meaning. Every sentence resolved to exactly one interpretation. The entire universe could be described by a state vector of about a hundred variables. This closure was not a limitation Winograd failed to notice—it was the experimental condition that made the demonstration possible. The blocks world was engineered so that the gap between symbol manipulation and genuine understanding could not be detected from the output.

SHRDLU integrated syntax, semantics, and reasoning more tightly than any previous system. It parsed sentences according to a grammar, mapped them onto semantic representations using rules linking syntactic categories to domain predicates, evaluated those representations against its model of the blocks world, and generated responses by reversing the process. At no point did anything happen that could be described, without metaphor, as understanding. The program did not know what a block was. It did not know what 'red' looked like. It operated entirely within a formal system that happened to use English words as tokens. Winograd himself would later state with characteristic precision: the program did not understand English—it understood the formal language of the blocks world, which happened to be written in English.

The AI community's response was euphoric. Natural language understanding—the hardest problem in AI, the capability that would distinguish genuine machine intelligence from mere calculation—appeared substantially solved. What remained was engineering: scaling up, expanding vocabulary, adding domains. The fundamental problem of getting a machine to understand what a human meant when speaking English had been cracked, and SHRDLU was the proof. Winograd's doctoral thesis, published in 1972 as 'Understanding Natural Language,' was received as a landmark. Few suspected, at the time, that the builder would spend the next fifty years explaining why the landmark demonstrated the opposite of what it appeared to demonstrate.

Origin

SHRDLU emerged from MIT's Artificial Intelligence Laboratory between 1968 and 1972, during the first wave of serious attempts to build machines that could process natural language. The name itself—SHRDLU—comes from the sequence of letters on a Linotype machine's second row, a piece of printing history repurposed as an acronym for nothing. Winograd was twenty-two when he began the project, working within the intellectual environment that Marvin Minsky and John McCarthy had established—a culture of extraordinary ambition, technical virtuosity, and shared conviction that intelligence was computation.

Key Ideas

Closed-world success. SHRDLU worked brilliantly within a domain so constrained that complexity, ambiguity, and context-dependence had been legislated out of existence—every attempt to extend its methods to broader domains failed.

The illusion of understanding. What looked like comprehension was formal symbol manipulation—the program parsed strings, mapped them to representations, and generated responses according to rules, with no experiential engagement with meaning.

Breakdown as revelation. The moments when SHRDLU encountered sentences outside its capabilities were the moments when the true nature of the system became visible—smooth performance concealed the walls, breakdown revealed them.

The indistinguishability problem. Within the blocks world, there was no observable difference between a system that understood English and a system that executed formal procedures using English words—the gap became visible only at the domain's boundaries.