SHRDLU, 1970: the blocks world where a machine first appeared to understand
At MIT’s Artificial Intelligence Laboratory in 1970, a graduate student sat at a CRT screen and typed at a program he had built: “Pick up a big red block.” The program replied: “OK.” He typed: “Grasp the pyramid.” It replied: “I DON’T UNDERSTAND WHICH PYRAMID YOU MEAN.” They negotiated. He asked how many blocks were not in the box. It answered: “FOUR OF THEM.” For the first time in artificial intelligence, a computer appeared not just to be responding but to be listening.
The program was called SHRDLU — a name borrowed from a printer’s fossil. On Linotype typesetting machines, the two leftmost columns of keys spelled out ETAOIN SHRDLU: the twelve most-common letters in English, arranged in frequency order. When an operator made a mistake mid-line, he’d sometimes drag his fingers down those columns and let the garbled slug go to press, leaving ETAOIN SHRDLU buried in newspaper galleys around the world as a typographic ghost. Terry Winograd, a 23-year-old PhD student studying natural language at MIT, adopted the back half as his program’s name. It was exactly the kind of quiet joke that announces you are not taking yourself too seriously.
Winograd built SHRDLU between 1968 and 1970 under the supervision of Seymour Papert — the same Papert who had just co-authored Perceptrons with Marvin Minsky. The program ran on a DEC PDP-6 in Lisp and a special-purpose language called Micro-Planner. Its world was carefully bounded: a virtual box containing colored blocks, pyramids, and a ball, rearranged by a simulated robotic arm on command. The vocabulary ran to about fifty words. The trick was not the vocabulary.
SHRDLU could track context across turns. If you said “put the red block on the green one” and then “move it to the right,” the program knew what “it” referred to. It could reason about physical constraints: if a command was impossible because something was in the way, SHRDLU said so and cleared the path first. It remembered its own prior actions. Published as Winograd’s MIT doctoral thesis in February 1971, the work arrived with a lengthy printed transcript of human-machine dialogue so fluent it looked like genuine comprehension. The Computer History Museum describes it as the first program to successfully combine syntax, meaning, and deductive reasoning for natural language understanding.
Here is what Winograd said in a 1991 interview: “The famous dialogue with SHRDLU where you could pick up a block — I very carefully worked through, line by line. If you sat down in front of it and asked a question that wasn’t in the dialogue, there was some probability it would answer it.” That probability was not high. The transcript had been curated. SHRDLU was an expert performer in a very small theater, and the theater turned out to be the point. When researchers tried to expand the blocks world into anything resembling the real world, the procedural framework collapsed. Natural language, it turned out, required not fifty words about colored blocks but an open-ended shared world — which is to say, everything.
What survived the collapse was a set of insights that still structure the field: that language understanding requires a model of the situation being described; that syntax and semantics cannot be separated; that context is not decoration but load-bearing. Winograd himself left AI for human-computer interaction, having concluded that the gap between appearing to understand and actually understanding might not close by the symbolic route.
The researchers who disagreed would spend three more decades finding out he was right about the hard part — before deciding to replace symbols with statistics, and the blocks world with the entire internet.
Sources
- SHRDLU — Wikipedia — technical architecture (DEC PDP-6, Lisp, Micro-Planner), the blocks world design, Winograd’s 1991 remarks on the curated transcript, and the AI community’s response.
- AI & Robotics Timeline — Computer History Museum — SHRDLU as the first program to combine syntax, meaning, and deductive reasoning to understand natural language commands.