Part of the lesson is that we have no idea how human minds work, or little enough idea that our models are quite disconnected from actual cognition. If our models were accurate enough, then we would not have to spend decades fishing for the right algorithms.
This has not been helped in recent decades. We use computer speak to describe human cognition, and the dual use of vocabulary fools generations of brilliant minds into thinking we are similar to these machines we command.
Another lesson is that progress in AI is likely to continue tracking closely with the availability of computational resources.
Rodney Brooks wrote a response here: http://rodneybrooks.com/a-better-lesson/