Terry Winograd Builds SHRDLU at MIT

The history of AI is a story of disillusionment. The reasons to get excited (and afraid) are easy enough to predict and see coming. The problem is time frame and so often theory and ideas had to wait (decades typically) for other fields to advance while their ideas were considered dead and disproven. History praises the true believers like Geoff Hinton who stuck with it through thick and thin. What about the field's apostates?

Perhaps no figure in this story represents disillusionment and renunciation as precisely as Terry Winograd.

His story overlaps with an early goal of the AI field: a machine that genuinely understood natural language the way people do. Why should humans spend so much time learning machine languages? Shouldn't they learn ours instead?

Well, the issue is there are different ways of going about this task and two tracks had already settled by early 1970s, learning and rules. But since the incredibly damaging blow to the learning side of Minsky's Perceptrons in the previous year, the rules side (symbolic, expert systems AI) was having a moment.

Emblematic of that moment for expert systems was Winograd's SHRDLU. Between 1968 and 1970, Terry Winograd wrote SHRDLU at the MIT Artificial Intelligence Laboratory. A user typed English on a teletype, while the program could "reply" in a very limited way. It was a simple robot with only a hand and an eye that could move blocks, pyramids and a box shown on a display screen. But that was about it. It was an expert system after all. Every outcome and scenario and answer had to be anticipated and baked in.

There was a lot of confusion. It asked which pyramid was meant when an order was unclear and even tried to account for its actions. Pressed on why it had stacked three blocks, it answered, “BECAUSE YOU ASKED ME TO.”

The name you ask? It spells the second column of keys on a Linotype typesetting machine.

Perhaps because of the clear and obvious limitation of expert systems, Winograd eventually abandoned the field entirely, rejecting it in a detailed and fundamental way. In a 1986 book with Fernando Flores, Understanding Computers and Cognition, Winograd and Flores argued that the task Winograd had spent decades on would be a failure even if it were to ever succeed. That's because, they argued, even fluent-seeming machines could never actually understand because they lacked embodiment and participation in life.