Critical Code Studies

Critical Code Studies (CCS) applies the interpretive methods of the humanities to computer source code, reading code, architecture, and documentation within their sociohistorical context. Its premise is that lines of code are not value-neutral: meaning grows out of what the code does but is never exhausted by it. Code can be close-read like any other text, attended to for its naming, its comments, its silences, and the assumptions it encodes, while still being held to its material test, that it must compile and run.

This reading works in the tradition developed by Mark C. Marino¹ and David M. Berry², reading code at once as language, mechanism, world-model, and social repository. SHRDLU is an unusually rich object for this method, because its central claim was itself a claim about meaning: that to represent knowledge as procedures was to let a machine understand.

• Mark C. Marino, "Critical Code Studies" · the founding essay (Electronic Book Review, 2006).
• Mark C. Marino, Critical Code Studies · the field's book-length statement (MIT Press, 2020).
• David M. Berry, "Reflections on Method for Critical Code Studies" · the method this project follows.

¹ Mark C. Marino, Critical Code Studies (Cambridge, MA: MIT Press, 2020), the field's book-length statement, following his founding essay of the same title (Electronic Book Review, 2006). ↩

² David M. Berry, The Philosophy of Software: Code and Mediation in the Digital Age (Basingstoke: Palgrave Macmillan, 2011); and, for the collaborative method applied to a contemporary of SHRDLU, David M. Berry and Mark C. Marino, "Reading ELIZA: Critical Code Studies in Action" (Electronic Book Review, 2024). ↩

Why read this program now

We are once again surrounded by machines that appear to understand language. The large language models of the 2020s answer in fluent prose across an open domain, where SHRDLU could converse only about a tabletop of blocks. Yet the two stand in an instructive symmetry. SHRDLU's understanding was grounded, every word tied to procedures over a world it could inspect and act on, and it could explain its reasoning step by step, but it could not leave its microworld. Today's models are general but, in the classic phrase, arguably ungrounded: fluent everywhere, anchored nowhere, and opaque as to why they say what they say.

Reading SHRDLU's code returns us to the first, most optimistic version of the dream, and to the moment its limits became visible. The history of that disappointment, the move from microworld to scaling problem to "AI winter", is also the prehistory of the present.

The object: a conversation about blocks

SHRDLU's world was a simulated tabletop, the paradigmatic "microworld" of the Minsky–Papert research programme: a box, a table, and a handful of blocks, cubes and pyramids in red, green and blue, moved by a simulated arm. The system was built from four parts, all in MacLisp under the ITS operating system: a parser; a grammar of English written as procedures (Winograd's PROGRAMMAR, after Halliday's systemic grammar); programs for semantic analysis; and a deductive problem-solver in Micro-Planner. The conversation here is the demonstration dialogue Winograd published, reproduced from his own transcript and running on those parts: the person types in lower case, SHRDLU answers in capitals.

Read the full demonstration dialogue, with commentary →

The machine and its world

That wager has lately returned. After the success of large language models and diffusion models, much of contemporary AI has turned to world models: the argument, pressed by Yann LeCun in his work on predictive world models (his JEPA programme) and by Fei-Fei Li under the name "spatial intelligence", that an agent which acts in the world needs an internal model of it, and not only a facility with text or images. SHRDLU is an early, miniature instance of exactly that idea, a program that kept and reasoned over a model of its world, which is why its microworld is worth reading closely now rather than only as history. The overview follows that thread to the philosophical question it raises.

The reading

The study works across the registers of Critical Code Studies, pairing close technical explication of a code fragment with the cultural constellation that spirals out from it.

1. A name that means nothing. ETAOIN SHRDLU: a pseudo-acronym drawn from Linotype letter-frequency columns by way of Mad magazine and, probably, a 1942 science-fiction story about an AI typesetter. A program about meaning, named with a deliberate non-word.
2. Procedures as representation. The thesis at the program's heart: that to represent knowledge as procedures is to let a machine understand. PROGRAMMAR, systemic grammar, and procedural semantics, read as a theory of meaning embedded in code.
3. The closed world. The blocks world as microworld: we look at what gets modelled, what is abstracted, and how the limited domain is important for enabling a world model to function.
4. The goal stack speaks. A machine that can answer "why?" because its reasoning is an explicit stack of goals. Self-explanation as designed effect, and its distance from understanding.
5. Demo or die. The scripted dialogue as artefact. Winograd's own later verdict, the "Potemkin village", and the microworld scaling problem that bridges to the AI winter.
6. Grounded but not general. SHRDLU read against the large language models: the symmetry of grounding and generality, and what the first dream of understanding can still tell us.

Contact

The SHRDLU reading is a Critical Code Studies project. Correspondence and corrections are welcome.