March 31, 2016
On The Digital Life this week, we chat about the evolution of artificial intelligence in light of recent public failure and success by tech giants in the AI space. First, Microsoft had to terminate Tay, its teenage chatbot, after the bot started tweeting neo-Nazi propaganda and other abusive language at people. Meanwhile, Google’s DeepMind has created an AI capable of beating some of the very best human players in the world at Go, the Asian strategy board game.
Microsoft Terminates Its Tay AI Chatbot after She Turns into a Nazi
In a Huge Breakthrough Google AI Beats a Top Player at the Game of Go
It sounds more like artificial intelligence runs slightly amock but it’s also worth reflecting on that in these areas of social interaction which are millions of times more complicated than rules based systems, that AI is probably going to have a long way to go. Dirk, do you want to share your thoughts on poor Tay and Microsoft’s falling on its face recently?
It’s just a recipe for disaster. Unless you program it, thinking two or three moves ahead to taking for granted that it’s going to be a lot of horrible, ugly, demagoguery that this bot is exposed to. How do you design it to deal with that proactively as opposed to reduce itself to that behavior. Just from a very conceptual level, Microsoft deploying it in this place, in this way, as sort of this innocent learner, coming through the world, really could anything have happened but it be twisted into a horrible, ugly, racist asshole, I don’t know.
The number of options, especially once you start looking ahead a few moves, as opposed to just reacting to the way the board is now. That artificial intelligence would never or would take a very long time before it was able to be a top human player and that happening in 2016 is, I would say, decades ahead of where people thought that would happen. However, I don’t think it’s anything about the game itself that allowed the AI idea to the point of beating human. Unlike Tay, which we talked about a minute ago, which is fairly basic and then learning in real time within the environment of Twitter, this AI was trained by professional Go players for months, going back to last year.
An earlier version of this program had less spectacular, but still impressive results against not as good of a player, the player who was beaten recently, as the fourth ranked player in the world. You’re talking of the best of the best. Maybe he was sixth ranked but really someone at the top of the pyramid. The player was played last year was maybe 400th best which is still really impressive like that guy’s brilliant and awesome, but there’s a big difference from from fourth to 400th. The player who was playing last year, after they played, they became part of the team and they have been playing almost nonstop and programming this AI nonstop to participate and play Go. My point is that they have been working for months on making this thing as good as it can be with a professional player there, telling the computer how to think, telling the computer what moves to make, giving it all of this context that are super and deep and granular level, in order to be able to go and be a top human player.
It’s much more nuanced in bringing that subject matter expert and I think there’s multiple involved but there’s this one fellow, very interesting and thoughtful fellow. I’m not trying to say their names because I’ll butcher the pronunciation as that ignorant American and I don’t want to do that, but seems like a very thoughtful and interesting fellow, he’s been in their programming this thing nonstop along with a lot of computer scientists and other experts. Many people have been collaborating to build this machine, to get it to the point where it could beat a top human in Go. That’s not to take anything away from the magnitude of the accomplishment, but it’s really talking about, when we think about AI what what is it? Is it something that’s learning on its own and adapting? Not really, at the level we’re talking about here, it’s more of something that has had tremendous amounts of knowledge put into it and coded into it that it’s able to then take this huge collection of stuff and apply it in order to win the game.
It’s not like a human being looking at each move in the same way, reacting, responding, it’s doing it in very, very different ways. Those are different ways that can create interesting things. I believe it was in the second game that they played, this more recently against the fourth best player in the world, the computer early in the game made a move that was so unexpected and so different than any of the patterns on how the best top professionals play Go, that the human opponent had to get up and go to the bathroom to take a break. It blew his mind that this computer and this contacts, that clearly was playing at a skilled level, would make this super unusual move.
I think a lot of media glommed on to that to say, “Well, here’s an example of higher intelligence. This is sort of the next level.” I think it’s a little bit more rudimentary than that. The machine is just able to compute many more possibilities than the human is able to compute and was able to say, “Okay.” This will be a little bit too inside baseball and may not make sense to people who aren’t familiar with Go, but I’m going to abandon this a little tactical fight in this part of the board and put a stone way over here to start setting up strength for a future fight in that part of the board at a time and in a way that would seem completely inappropriate for what the situation is if I really know what the hell I’m doing.
We know that the computer has so much intelligence put into it, that it does have a very good idea what it’s doing. For that top professional to be challenged with that, that some entity, whether it be human or computer that knows what it’s doing, that clearly knows what it’s doing, makes this move that’s different than anything, that any top professionals ever contemplated, it becomes a very compelling moment but it’s a moment that’s the product of computation, it’s not the product of sentience, which I think is where some of that conversation went. Maybe I have been a little too rambling here and I might recap.
What would be easy, what would be great is if it knew when we say Thai restaurant, it’s talking about one specific place. If we say the Burritos, it’s talking about one specific place. Suddenly, that thing with the local contact as a bot will be able to make a round trip orders for us in seconds. We type in what we want and then it’s all just done, but with the bots being designed as these global things, it can’t discern when I say Thai, that it means the restaurant to our left, not across the street into our right. If I say pizza, it means the place that’s a lot farther, that’s really good as opposed to the place that’s closer and sucks.
There’s this gap with programming being done at the global level, we’re trying to write a bot that covers everyone everywhere in a generic way. The usefulness doesn’t get that deep. Where it gets deep is where it’s more localized, where it’s more specific to us in the examples I’m using just because of the very micro geography that we’re in and our taste preference as opposed to the more generic glom and so bridging that gap, is where a lot of the really exciting things will start to happen in the sort of software you’re talking about to really get at the local personal level and convert that into care as opposed to the generic and the macro.