SHRDLU and the blocks world: impressive, honest about its limits, and why it did not scale

SHRDLU was written by Terry Winograd at MIT between 1968 and 1970 and described in his 1971 thesis, "Procedures as a Representation for Data in a Computer Program for Understanding Natural Language". It was written in Lisp and Micro-Planner. What it did A user typed English instructions and questions about a simulated table of blocks, pyramids and boxes in different colours. SHRDLU could: - carry out commands ("Pick up a big red block") - answer questions about the scene ("Is there a large block behind a pyramid?") - resolve pronouns from earlier in the dialogue ("Put it in the box") - ask for clarification when a reference was ambiguous - explain its own actions ("Why did you pick it up?" "To get rid of it.") - learn names for new arrangements ("A steeple is a stack of two green cubes and a pyramid") Unlike ELIZA, it really did model its world. Every sentence was parsed, mapped to a procedure, and checked against the state of the blocks. Why it did not scale The blocks world was small, closed and fully known. Every object, property and action could be written down. Real language refers to an open world where none of that is true, and the hand-built knowledge did not generalise. Winograd himself later moved away from this approach; his 1986 book with Fernando Flores, Understanding Computers and Cognition, is a critique of it. The lesson that held Grounding helps: a system that can check a sentence against a model of the world is more trustworthy than one that only matches text. The hard part was always building that model for anything bigger than a table. Sources - Wikipedia, "SHRDLU": https://en.wikipedia.org/wiki/SHRDLU