Meet SHRDLU (1968-1970)

Long before large language models like GPT, there was SHRDLU. Developed by Terry Winograd at MIT, it was a groundbreaking experiment in "Symbolic AI" (also known as GOFAI: Good Old-Fashioned AI).

1. The "Micro-World" Hypothesis

Instead of trying to understand the entire universe (which is messy), SHRDLU lived in a restricted Blocks World. This world contained only specific objects: bricks, pyramids, and a box.

• Symbol Grounding: Because the world was small, every word ("red", "move", "on") could be perfectly defined. The program "understood" a block because it had a direct pointer to that object in memory.

2. Capabilities

• Understanding: "Put the red block on the green block."
• Reasoning: "I can't, the green block is covered." (It understood physics constraints).
• Memory: "Why did you do that?" -> "Because you asked me to."

3. Historical Context

It ran on a DEC PDP-10 mainframe (1 MIPS speed) and used a vector graphics display. The logic was written in MacLisp and Micro-Planner.