I fed GPT-4 some really old fashioned spatial reasoning questions (inspired on SHRDLU), which it passed. Then when questioned about unstable configurations (whi

I fed GPT-4 some really old fashioned spatial reasoning questions (inspired on SHRDLU), which it passed. Then when questioned about unstable configurations (which IIRC SHRDLU could not handle) it passed those too. So it seems like it is definitely capable of some forms of reasoning. Possibly we both tested it in different ways, and some forms of reasoning are harder for it than others?