What SHRDLU Knew, and What It Didn't

**EDO SEGAL:** I want to start this round with a confession, because the best questions I know come out of wounds. I started in Assembler. I was raised by machine code, and for the entire history of computing, using a machine meant translation — you compressed your intention into the machine's grammar and paid a tax on every conversion. In December 2025 I watched that tax go to zero. I stood in a room in Trivandrum with twenty engineers and watched each become capable of more than all of them together, because for the first time the machine met them in *their* language — mess, half-finished sentences, implication and all. Terry, you built the very first version of that feeling, in a tabletop of blocks, fifty-five years before my room in Trivandrum. So take me back. Tell me what SHRDLU actually knew. Slowly, the way you would tell a sharp fifteen-year-old.

**WINOGRAD:** Happily, because the lesson is in the details. SHRDLU knew that a green cone existed because I had asserted a green cone into its database. It knew that "the box" referred to a particular object because the symbol BOX was bound to that object in a table I wrote. Its understanding of "support" was a logical predicate — true or false — with no feel for the heft of a thing, no sense of how a tower wobbles before it falls. The meaning of every word bottomed out in a definition a human supplied. There was no point anywhere in the system where its grasp of "red" or "on" or "pick up" was answerable to anything outside the formal system itself. The symbols referred to other symbols, and the whole closed loop touched the world at exactly one moment: when a human being read the output and supplied, from their own life, the meaning the machine lacked. When SHRDLU printed "OK" after stacking a block, *I* understood that a block had been stacked — richly, with all my embodied knowledge of blocks and stacking. The machine had updated a data structure and drawn a picture. The illusion of shared understanding was produced by my competence, not its.

**HEIDEGGER:** And note — I must interrupt to mark it — that what Herr Winograd is describing is precisely the broken hammer, only it never had the unbroken state to fall from. Your blocks world was made entirely of the present-at-hand. Objects with properties, laid out, surveyed, calculated. There was no [absorbed coping](https://www.youonai.ai/fieldguide/med/absorbed_coping) beneath it, because there was no involvement for it to be the coping *of*. You did not build a defective version of understanding. You built a perfect specimen of understanding's opposite — the detached, representing stance, in pure culture, with the involved background that gives it meaning stripped entirely away. And then you were surprised it had no world. It could not have had a world. You had built the exception with no rule under it.

**WINOGRAD:** That is right, and here is the part that should keep your audience awake, Edo, because it is the bridge from my tabletop to your Trivandrum room. The blocks world's central limitation has not vanished in the modern machine. It has been hidden by scale. A large language model was not given a hand-written microworld — that is a real and important difference, and I will not let it be waved away. But it was given a finite corpus and a fixed architecture, and its world is the statistical structure of that corpus. It, too, has no involvement in anything outside its own formalization. The only difference is that its formalization is so vast — effectively the textual output of a civilization — that the walls are nowhere visible. SHRDLU's microworld was small enough that you could feel the edge in a sentence or two. The modern model's world is large enough that you can converse for hours and never strike a wall. But the edges are still there. They show up as the confident fabrication, the subtle incoherence under sustained pressure, the failure that arrives not gradually but as a sudden collapse when the conversation wanders into genuine novelty the training distribution never covered. The blocks world grew until it swallowed the visible horizon. It did not become the world.

**EDO SEGAL:** Let me restate that, because it is the engine of your whole position and I want to hold it where the reader can see it. You are saying — literally — that the difference between SHRDLU and the machine I watched amaze twenty engineers is *only* the size of the tabletop. That my Trivandrum machine is the blocks world grown so large nobody in the room can find the wall, but the wall is there, and it is the same wall. Is that the version you find most compelling, or is it too strong?

**WINOGRAD:** It is close to the version I will defend, with one honest correction. It is not *only* size, because at sufficient size something genuinely new appears in the behavior, and I will not pretend otherwise — I was surprised by it, and I will say so in a later round, because the surprise is the most useful confession I have. But it is *structurally* the same wall. SHRDLU had a foreground — the pre-formed representations I gave it — and no background. The modern model absorbed, from an ocean of human text, something that *functions* behaviorally like a background. That is the new thing, and it is why my old categories no longer fit cleanly. But function is not the same as having. My claim, the one I will hold all night, is that the machine has an extraordinary model of the *traces* of understanding, and that the involvement those traces are traces *of* is still, exactly as in 1970, on the human side. Larger. Aggregated across billions of texts. Distilled into weights. But ours.

**HEIDEGGER:** Here I want to press my colleague, because he has conceded a hair more than I would, and the hair matters. You say it functions like a background. I say a background is not a function. The background — what I called the unarticulated whole within which anything can show up as relevant — is not a thing the system *has*, the way it has weights. It is the *being* of the entity for whom things matter. The carpenter's background is not stored in him; it *is* him, his thrownness, his having-been-raised among hammers and wood and the company of other carpenters, his mortality giving the work its urgency. You cannot scrape that off a corpus, because it was never in the corpus. It is the condition under which the corpus was written. The machine has the wake of a thousand backgrounds and no background of its own, and a wake is not a smaller boat. It is the absence of a boat, given shape by water.

**WINOGRAD:** I agree with that, and watch how carefully I have to agree now, because here is where the engineer in me will not let the philosopher have the whole victory. There is a second lesson in SHRDLU, and it is about the danger of success, and it indicts *us*, not the machine. SHRDLU was misleading precisely because it worked so well within its walls. Observers extrapolated to a future where the walls would simply expand to encompass everything, and that extrapolation was the engine of the optimism before the first AI winter, and it was wrong — because the difficulty of the real world is not that it is large but that it is *open*. It contains the unanticipated as a permanent feature, not as a gap you fill in. We are running the identical error now, on a grander scale, and the grandeur is what makes it lethal. Today's systems are far more impressive than SHRDLU across far more domains, which makes the extrapolation far more seductive. So I will give Heidegger his background and take something back: even if I am *wrong* about the metaphysics, even if there is some strange new kind of background in those weights, the honest reading of SHRDLU's history is still a warning against reading present competence as a promise of future comprehension. The two came apart once, spectacularly. I am telling you, with my hands, that they can come apart again.

**EDO SEGAL:** What SHRDLU never knew, you wrote, was that it was in a blocks world at all. It had no way to register the boundary between what it could handle and what lay beyond — no experience of its own limits, because it had no involvement in anything that could push back. Hold that thread, both of you, because it returns when we get to the place Heidegger calls breakdown and the place the market calls the death cross. But first I want to follow the machine out of the tabletop and into the question that made Terry walk away. Not "can it produce the behavior of understanding" — it plainly can — but "where does understanding come from, and what does it require." That is the book you wrote with Flores, and it told the field, at the height of its confidence, that it had misunderstood understanding itself. After this.