Terry Winograd's SHRDLU (1968-1970) demonstrated groundbreaking natural language understanding through symbolic reasoning about a blocks world. This note examines SHRDLU's famous "support supports" sentence—a grammatically correct but cognitively challenging query—and observes how modern language models handle the same problem through fundamentally different mechanisms.
Terry Winograd's SHRDLU (1968-1970) demonstrated groundbreaking natural language understanding through symbolic reasoning about a blocks world. Running on a PDP-6 with less processing power than a modern microwave oven, it achieved capabilities that seemed almost magical.
The system used Micro Planner, a goal-oriented planner that could backtrack through logical deductions, maintaining explicit symbolic representations of objects and their relationships. SHRDLU could:
SHRDLU's most famous demonstration of linguistic prowess was parsing this question:
The sentence requires tracking nested relationships:
SHRDLU correctly answered "YES, THE GREEN PYRAMID" by systematically unwinding these dependencies through logical inference.
The following 2D diagram shows a valid blocks world where SHRDLU's answer is correct. The blue cube supports both the tall red pyramid and the small red cube. The red cube (the shortest thing the blue cube supports) does indeed support the green pyramid.
When asked to solve this problem, a modern LLM (Claude) successfully created a valid world configuration and visualization—but through entirely different means:
The LLM essentially "imagined" a world where the sentence would be true, then drew it—solving the problem through learned patterns rather than logical reasoning.
SHRDLU's approach—explicit symbolic reasoning with grammatical understanding—remains unintuitive to modern AI practitioners raised on neural approaches. Yet both systems reach correct answers. This contrast highlights how contemporary AI has achieved impressive capabilities while bypassing the structured reasoning that once seemed essential to intelligence.
The "support supports" sentence stands as a reminder that there are multiple paths to understanding, even if we've largely abandoned the road SHRDLU traveled.