SHRDLU understands a world of blocks

"Pick up a big red block." "OK." "Grasp the pyramid." "I don't understand which pyramid you mean." "Find a block which is taller than the one you are holding and put it into the box." "By 'it', I assume you mean the block which is taller than the one I am holding. OK." The transcript of Terry Winograd's SHRDLU, from his 1971 MIT thesis, still reads like a conversation with something that understands. The program parsed English grammar, kept a model of its world of blocks, planned actions in it, and could explain why it had done what it did.

The trick, and Winograd was clear that it was one, was the world. Everything SHRDLU could be asked about was on one simulated table, so its grammar, its knowledge and its reasoning could all be complete. Scaling the approach to a kitchen, let alone the world, meant writing down everything, and nobody could.

SHRDLU is the peak of the idea that understanding language is a matter of grammar plus a hand-built model of what is being talked about. Winograd himself concluded it would not scale and moved into human–computer interaction. The problem was eventually solved from the other end, by models that learned from text what SHRDLU had been told about blocks.