Two articles caught my attention today when I woke up. One was the list of 2017 Hugo Award winners (we’ll talk about that some other time). The other was an article from the Verge:
The world’s best Dota 2 players just got destroyed by a killer AI from Elon Musk’s startup
Sottek doesn’t go into much detail, so let me add some context. The OpenAi bot beat SumaiL, Arteezy, Pajkatt, Fogged and CCnC, and then took on Dendi in a limited 1-1 matchup. These are almost universally recognized as the best players in Dota2 right now. They’re smart, their reaction times are insane, and they’re pros at a game that even Musk acknowledges is tougher than chess or Go.
This is one of those landmark moments. Kasparov being beaten by Deep Blue. Lee Sidol being beaten by AlphaGo. Add this to the list. Granted, it was a limited matchup, but this is important, because the OpenAI bot did all of this with just two weeks of training.
This is an interesting area for me right now. I’m writing a short story for an anthology called the Expanding Universe, and I’m trying to imagine AI-controlled starships at war with humans. Naturally, I’ve been looking up the Kasparov and Sidol games.
There’s something that sticks very strongly in the mind. Fan Hui, the first human master to challenge AlphaGo, watches the game between Lee Sidol and the Ai. At the 37th move, he remarks: “It’s not a human move. I’ve never seen a human play this move. So beautiful. So beautiful.”
“It feels a little bit like a human, but at the same time it’s something else,” says the Dota2 player Dendi recently, talking about OpenAI’s bot.
What interests me is how similar the feedback of the Dota2 players is to that of the defeated Go champions. The Ai played (and I paraphrase) like a human . . . but not like a human.
The more I read, and the more I understand, the more I realize that a human can’t accurately imagine what the AI would do. While a battle scenario and the reactions on both sides might seem completely to us humans, to AI it would be a completely different thing. After all, these are not minds that have to deal with stress, or anxiety, or adrenaline, or all of the things we humans have to deal with in competitive situations.
Surface verdict: classic Terminator 2 – AI versus humans, especially in warfare, is going to go very badly. For the humans.
But AI versus humans is going to be an unlikely scenario. AI is a tool. We don’t ditch our tools that easily.
Consider chess, which has had a long time to adapt to computing. Today, the best chess players are not Ai, nor humans, but human-Ai hybrids – flesh-and-blood players who run a combination of chess engines in the background and select plays out of their calculations for the best possible moves. The field is called Advanced Chess.
It’s really not hard to imagine this use case naturally extending to every other thing where there’s a vital advantage to winning. I don’t think we’ll ever get to a point where there’s one Skynet on the other side and nothing on this side: I think we’ll end up with various AI-plus-human combinations versus each other – that might even be a more potent mix.
It’s still right for Elon Musk et al to be cautious of AI. If we’re heading into a future where human-AI combinations rule the roost, a a lot depends on what kind of human you have with the AI. Pick the wrong human, and that’s like giving a monkey the key to the banana plantation.
What next? I believe the next game / frontier will be the god of all complex multiplayer games – Starcraft II. Google Deepmind is already being trained to take on the SCII scene.
Once Deepmind masters Starcraft, we will need to terms with the fact that Ai are better than we humans will ever be at competition – specifically, at strategy and war. They will think in ways we never thought possible, and they will win. I wonder how long it’ll be before a general sits with his advisors, who are all AI of a screen, and all the human has to do is pick the best possible option.