An overview of SHRDLU

The return of the world model

It is worth reading SHRDLU now, rather than only as history, because contemporary AI has circled back to the question it posed. After the success of large language models and of diffusion models, much of the field has turned to the idea of the world model: the claim that a system which only predicts the next token, or denoises an image, does not thereby possess a model of the world it could plan and act within, and that intelligence worth the name requires one. Yann LeCun has made this his central thesis, arguing that autoregressive language models are a detour and that an agent must learn a predictive model of the world, the programme he calls JEPA; Fei-Fei Li has named the same gap "spatial intelligence" and founded World Labs to pursue it. The shared wager is that agentic AI, systems that act in an environment rather than merely answer questions about it, cannot do without some model of the world they act in.

SHRDLU is the prehistory of that turn, and also its oldest objection. Its blocks world was already a world model, hand-built and fully inspectable, over which the program planned, acted and explained itself, the very capacities now being sought at scale. What SHRDLU also makes plain is how much had to be excluded to make such a world tractable, and this is where the philosophical critique it provoked still bites. Drawing on Heidegger, Hubert Dreyfus argued that artificial intelligence succeeds only inside artificially bounded micro-worlds because it has no world in the deeper sense. In Heidegger's vocabulary such a system is world-poor (weltarm): able to register objects and relations, but not to dwell in a meaningful whole the way a being that is in-the-world does. A learned world model far larger than SHRDLU's is still a model of a world held at arm's length, not a being-in-the-world, and the question Dreyfus pressed, whether understanding can be had without embodiment, situation and care, is precisely the one the world-models turn must now answer. Read this way SHRDLU is not a quaint precursor but the place where the stakes were first made clear: the demand that an intelligent agent have a world, and the lasting difficulty of saying what having a world would mean.

How it worked

SHRDLU was not a single trick but an integration of four cooperating components, each addressing a different layer of the problem of taking a typed English sentence and turning it into an action or an answer. What made the system persuasive was that these layers spoke to one another, so that syntax, meaning, and reasoning were brought into a single loop rather than handled in isolation. Winograd's central thesis, captured in his dissertation title, was that procedures themselves could serve as the representation of linguistic knowledge.

What it could do

Within its small domain SHRDLU did things that no earlier program had managed to combine. It resolved reference and anaphora in context, so that a pronoun such as "it" or a phrase such as "the green one" could be understood against what had just been said and what currently stood on the table. It queried its own model of the world, reporting on which objects existed and how they were arranged. It could answer "why" questions by reading back its goal stack, reconstructing the chain of intentions behind an action it had taken. It learned new words when told what they meant, accepting a definition for "steeple" and using it thereafter. And it refused requests it judged impossible, declining to perform what its model of the world would not allow.

The exchange reads as a small marvel of self-explanation, the program tracing its actions back through its own reasoning to the instruction that prompted them. You can read the full dialogue, and follow how the conversation builds context turn by turn.

The contrast with ELIZA, Joseph Weizenbaum's 1966 chatbot, is instructive. ELIZA produced its uncanny effect with almost no machinery: it matched keywords in the user's input and reflected them back as questions, keeping no model of the world and no memory of what had been said, so that its apparent understanding was an effect produced in the reader rather than in the program. SHRDLU works the other way round. It parses the sentence, consults and updates an explicit model of its blocks world, reasons about what is possible there, and can give an account of its own actions afterwards. Where ELIZA simulated understanding by deflecting it, SHRDLU tried to earn it by building a small world to be understood. On ELIZA and what it reveals about the wish to be understood by a machine, see the team's Inventing ELIZA (MIT Press, 2026).

Why it mattered, and what came next

SHRDLU is rightly remembered as the high-water mark of good old-fashioned AI (GOFAI). It seemed to demonstrate that symbolic representation and deductive reasoning could yield genuine linguistic understanding, and the optimism it produced ran well ahead of what the program could in fact do. That gap is the subject of the critique and legacy page, which develops three lines: the scripted, "demo or die" character of the famous dialogue, worked through line by line; the microworld scaling problem, the difficulty of extending an intelligence purchased by radically restricting the world; and Winograd's own later turn, from builder to critic, culminating in Understanding Computers and Cognition with Fernando Flores in 1986.

For more information about how the program got its name, see the name, and for the people involved in building it, see people.