SHRDLU, understanding, anthropomorphisation and hindsight bias

EDIT: Since I didn’t make it suffi­ciently clear, the point of this post was to illus­trate how the GOFAI peo­ple could have got so much wrong and yet still be con­fi­dent in their be­liefs, by look­ing at what the re­sults of one ex­per­i­ment—SHRDLU—must have felt like to those de­vel­op­ers at the time. The post is par­tially to help avoid hind­sight bias: it was not ob­vi­ous that they were go­ing wrong at the time.

SHRDLU was an early nat­u­ral lan­guage un­der­stand­ing com­puter pro­gram, de­vel­oped by Terry Wino­grad at MIT in 1968–1970. It was a pro­gram that moved ob­jects in a simu­lated world and could re­spond to in­struc­tions on how to do so. It caused great op­ti­mism in AI re­search, giv­ing the im­pres­sion that a solu­tion to nat­u­ral lan­guage pars­ing and un­der­stand­ing were just around the cor­ner. Sym­bolic ma­nipu­la­tion seemed poised to fi­nally de­liver a proper AI.

Be­fore dis­miss­ing this con­fi­dence as hope­lessly naive (which it wasn’t) and com­pletely in­cor­rect (which it was), take a look at some of the out­put that SHRDLU pro­duced, when in­structed by some­one to act within its simu­lated world:

Per­son: Pick up a big red block.
Com­puter: OK.
Per­son: Grasp the pyra­mid.
Com­puter: I don’t un­der­stand which pyra­mid you mean.
Per­son (chang­ing their mind): Find a block which is taller than the one you are hold­ing and put it into the box.
Com­puter: By “it”, I as­sume you mean the block which is taller than the one I am hold­ing.
Com­puter: OK.
Per­son: What does the box con­tain?
Com­puter: The blue pyra­mid and the blue block.
Per­son: What is the pyra­mid sup­ported by?
Com­puter: The box.
Per­son: How many blocks are not in the box?
Com­puter: Four of them.
Per­son: Is at least one of them nar­rower than the one which I told you to pick up?
Com­puter: Yes, the red cube.

It’s hard not to pro­ject an in­tel­li­gence into those re­sponses. Hu­mans are the main pro­duc­ers of lan­guage, so we’re prob­a­bly primed to an­thro­po­mor­phise any­thing that speaks to us this way; add to that SHRDLU’s seem­ing abil­ity to cope with some level of am­bi­guity (“Is at least one of them...”—one of what?), its very hu­man turns of phrase (“By “it”, I as­sume you mean...”, “I don’t un­der­stand...”) and you can see why naive out­siders could be­lieve that SHRDLU might show gen­uine un­der­stand­ing. It would be nat­u­ral to give SHRDLU the benefit of the doubt, and as­sume it was far more com­plex that it re­ally was.

We can also see why smart in­sid­ers might also be­lieve that. De­spite Dreyfuss’s cri­tique of AI, many AI re­searchers hadn’t yet grasped how badly sym­bolic ma­nipu­la­tion sys­tems would fail in real-world am­bigu­ous situ­a­tions. To them, SHRDLU’s perfor­mance would seem as con­fir­ma­tion, not that SHRDLU was very com­pli­cated (since they knew well how com­plex it was), but that un­der­stand­ing wasn’t that com­pli­cated (since SHRDLU seemed to demon­strate that).

I would posit that this wasn’t an un­rea­son­able be­lief at the time. They had a product—AI—that had demon­strated high-seem­ing perfor­mance in con­trol­led tests (a proof of con­cept, if you will), and that they were hop­ing to de­velop for more gen­eral us­age. In­deed, how do we know that SHRDLU failed to show some form of true un­der­stand­ing? Mainly be­cause the ap­proach failed in more gen­eral situ­a­tions. Had sym­bolic AI gone on and suc­ceeded in pass­ing Tur­ing tests, for ex­am­ple, then we prob­a­bly would have con­cluded that “SHRDLU was an early ex­am­ple of AIs with un­der­stand­ing.”

But of course, at the time, re­searchers didn’t have the cru­cial in­for­ma­tion “your field will soon be­come a dead-end where gen­uine AI is con­cerned.” So their be­lief in SHRDLU and in the whole sym­bolic ap­proach was not un­rea­son­able—though, like ev­ery­one, they were over­con­fi­dent.