> SHRDLU_

In 1970, Terry Winograd, a 24-year-old PhD student at MIT, demonstrated a program called SHRDLU that seemed to achieve something remarkable: genuine natural language understanding. SHRDLU inhabited a virtual "blocks world" — a simulated tabletop containing colored blocks, pyramids, and boxes that could be stacked, moved, and rearranged. Users could type commands and questions in plain English, and SHRDLU would understand them, execute actions, and respond intelligently. You could type "Pick up a big red block" and SHRDLU would identify which block you meant, check if it was clear, and move it. You could ask "Why did you put the blue block on top of the green one?" and SHRDLU would explain its reasoning, referencing earlier commands in the conversation.

The dialogues SHRDLU could handle were genuinely impressive. It understood pronouns ("Put it in the box"), resolved ambiguities ("Which cube do you mean? The large one or the small one?"), and maintained context across dozens of exchanges. It could answer questions about what it had done, why it had done it, and what it knew about the current state of the world. For researchers who had been struggling with the apparent impossibility of natural language understanding, SHRDLU seemed like a breakthrough. The system's name, incidentally, came from ETAOIN SHRDLU — the order of letter frequency on a Linotype machine — a playful reference that reflected the youthful exuberance of the MIT AI Lab in that era.

But SHRDLU's success concealed a fundamental problem that would haunt AI for decades: the blocks world was absurdly simple. It contained a handful of objects with a handful of properties, governed by straightforward physical rules. There was no ambiguity about what "red" meant, no uncertainty about whether one block was "on" another, no background knowledge required about culture, physics, emotions, or common sense. When researchers tried to extend SHRDLU's approach to richer domains — even slightly richer ones — the system collapsed. Every new concept required hand-coded rules, and the number of rules needed grew exponentially with the complexity of the world being modeled. This was the "frame problem" in action: how do you represent everything a system needs to know about the world, including all the things that don't change when something happens?

Winograd himself came to view SHRDLU as a cautionary tale rather than a triumph. He gradually moved away from AI research and toward human-computer interaction, arguing that the goal of fully automated understanding was misguided. He became a professor at Stanford, where he co-founded the influential "Understanding Computers and Cognition" framework with Fernando Flores. But perhaps Winograd's most consequential contribution at Stanford had nothing to do with natural language processing at all: in the late 1990s, he supervised the PhD research of a young student named Larry Page, who was working on a system for ranking web pages by their link structure. That system became Google. The man who had built the most impressive language understanding system of the 1970s helped launch the company that would, decades later, build the most impressive language understanding systems of the 2020s.