AlphaStar and the fairness of fast hands
DeepMind's StarCraft agent beat two pros 10-0, then lost the one game where it had to move a camera like a person. What counts as intelligence when the body is different.
DeepMind showed AlphaStar on Thursday. It’s a StarCraft II agent, and in recorded matches from December it beat two professional players, TLO and MaNa, five games to zero each. Then in a live exhibition game on the stream, MaNa beat it.
StarCraft is a much harder game for AI than Go. You can’t see the whole map, the game runs in real time, there are hundreds of unit types and thousands of possible actions at any moment, and a single match lasts tens of thousands of steps. AlphaStar was trained first by imitating human replays, then in a league of agents playing each other for what DeepMind says is the equivalent of up to 200 years of games per agent. The strategies that came out of the league are strong and sometimes strange, which is familiar by now from AlphaGo.
The discussion afterwards, especially among StarCraft players, has been about something else: the hands. DeepMind capped AlphaStar’s average actions per minute at a level comparable to pros, around 280. But averages hide bursts. In some fights AlphaStar’s rate spiked far above what humans can sustain, and those actions were perfectly precise. It could microcontrol individual units in three places at once with no misclicks. Human pros hit high APM numbers too, but much of it is spam clicks that do nothing. Every one of AlphaStar’s actions counted.
The other issue was the camera. In the ten recorded games, AlphaStar saw the whole map at once through a raw interface, where a human has to move a camera around to see what’s happening. For the live game DeepMind used a new version that had to control a camera like a person does, and that’s the one that lost, partly because MaNa found a way to confuse it by harassing its base.
I don’t think this makes the result less impressive. It does show something I find important. We compare AI to people as if the only difference is in the brain, but a lot of what makes human play human is the body: one pair of eyes that has to look somewhere, hands that take time to move and make mistakes, reaction times, fatigue. If you remove those limits, the agent can win with speed and precision in places where a human has to win with judgment. That’s a real skill, but it’s a different contest.
If what you want from these games is to learn about strategy and reasoning, the agent should have to live inside roughly human limits: a camera, a limited click rate with some noise, reaction delays. And if you want agents that eventually operate in the physical world, those limits aren’t a handicap to argue about. They’re the job. A robot’s joints have torque limits and its cameras have a field of view. An agent that only wins with inhuman bandwidth hasn’t learned the thing we need it to learn.