What does it mean to understand? This seems like a simple question but it isn’t. These days, AI is a field that attempts to understand simple English sentences like “Alexa play me a Beatle’s song.” But the understanding required to perform this command is considerably different than what is needed to read A Tale of Two Cities and understand it. Trying telling Alexa that “it was the best of times and it was the worst of times.” Say that to a person and they might say: “what do you mean?”
Or, consider this from today’s New York Times:
President Trump, ramping up his assertions of extraordinary powers, declared in a tweet that he had “the absolute right” to pardon himself for any crime.
The claim was the latest in an aggressive series of moves, as Mr. Trump and his legal advisers have argued that his authority places him above the reach of obstruction statutes.
What does a person do when they read this? They have to rely upon knowledge about the situation that they had before they read this. You have to know, for example, about Mueller’s investigation of Trump and you have to know about Presidential pardons. You also have to have the ability to realize that no one worries about pardons unless they are afraid of being found guilty of a crime. So, an intelligent reader who realize that this is discussion is based on the idea that the reader, and Trumps’ legal advisors, must have been thinking about what happens if the President is indicted. We have nothing like that in modern AI, no could we, because counting words won’t help. Why not?
Because the word “indicted” doesn't appear anywhere in the Times story. How would Alexa or Siri understand this article? They wouldn’t. Any intelligent reader knows that indictment is the underlying issue. And what would modern AI systems be able to do with this? Nothing. Modern AI systems are not actually about understanding no matter what IBM says about “cognitive computing"
The imminent arrival of AI is in everything we read these days due to a serious misunderstanding of the real issues in AI, which haven’t changed over the last 50 years.
Old time AI folks like me were always concerned with what it means for a human to understand.
Yesterday, there was a interesting article in the Boston Globe about an old AI guy (Patrick Winston) at MIT who is trying to address what understanding really is:
It won’t get much attention because what Winston is doing is not what the AI hype is all about.
Hi,
ReplyDeleteThis is the second article of yours along these lines i am reading. And as someone who is sincerely interested in actual cognition. Where would you advice i get involved or get started on a path to seeing its fruition, or what in your opinion is the correct direction, is there an approach that isn't currently being given the light of day i can look at instead of all the ML methods out there? If you have something written , you could quote in your response and i'd be glad to read. Thanks alot.