5 ms·
This isn't something AlphaGo was trained to do - that is, whether or not a board state it has seen was a human or a non-human move was not part of its training
by AaronFriel 11y ago
This isn't something AlphaGo was trained to do - that is, whether or not a board state it has seen was a human or a non-human move was not part of its training model.
To take this to its fullest conclusion, imagine that AlphaGo was not only trained on that information, that is, and that in addition the hundreds of thousands of board states it has seen, it also had layers which encoded analysis of the players before, during, and after the game (their tweets, their weibo messages, and so on). This is the difference between what AlphaGo does and what a human player can do, and what a true strong AI with unlimited compute power could do. AlphaGo can't play as efficiently as possible, it doesn't have that information.
A hypothetically omniscient being, or the Hand of God, can play meta-go. It can play in a way that cause the human player to be less likely to play a winning game. As Eliezer Yudkowsky put it:
With regards to tonight's match of Deepmind vs. Sedol, an example of an
outcome that would indicate strong general AI progress would be if a
sweating, nervous Sedol resigns on his first move, or if a bizarre-seeming
pattern of Go stones causes Sedol to have a seizure.
The Hand of God could play the game in such a way that, on games 1-3, it ended the game in a board state that issued a threat to Lee Sedol's family.
Fortunately it seems we're still a ways off from playing Go against the Hand of God.
- CamperBob2 11y agoWhen Lee got up and left the table for several minutes during the second game, I actually wondered if they were going to find him in the men's restroom in the lobby with his wrists slashed. It was genuinely kind of spooky when the camera kept returning to his empty chair, and it was a relief when he came back.
- yongjik 11y ago> ... or if a bizarre-seeming pattern of Go stones causes Sedol to have a seizure. I didn't think I could have a lower opinion of Yudkowsky, but apparently I was wrong. What is this, a Gibson fan fiction?
- AaronFriel 11y agoI don't think his comment was to be taken entirely seriously, but why do you have such a low opinion of him?
- marvin 11y agoI really don't get the hate against Yudkowsky that I've seen on HN recently. Disagreement is one thing, but it almost seems as if a lot of people here find his ideas and writings insulting. I don't get it; Yudkowsky is a philosopher and technologist with very interesting thoughts about an important subject that is not studied enough. It would be nice if people could substantiate their criticism rather than resort to name-calling. Everything I've read about Yudkowsky and MIRI seems dead-on: Further research around the issue of long-term AI safety sooner rather than later. I really don't get why this is problematic; I think it's a very good thing that someone smart spends their effort on this. The comment in question was obviously meant tongue-in-cheek to illustrate a point about potential more efficient, hypothetical ways of winning a game.
- yongjik 11y agoI consider his theories as the modern-day equivalent of "How many angels can dance on the head of a pin," or (according to some stories I've read) how 19th century folks worried that large cities would be buried in horse manure. YMMV, of course.
- NamTaf 11y agoIt's just his particular way of phrasing the concept of playing against the opponent and the game. Sort of like with poker, whereby you play not only against the cards on the table but also the minds of your opponents, knowing what they might do given a partiuclar situation.