Despite what people in the 90s thought, computers have vastly improved the chess experience. We now have an objective arbiter that can tell us when a move is really better. It makes things much, much more interesting than it used to be and makes the game accessible to amateurs such as myself.
It’s like everyone having their own personal chess commentator telling you what is going on. If you don’t play chess, I would say it would be similar to having a seasoned quarterback (John Elway for example) explain what is going on during a football match.
AlphaGo will do the same thing for Go. It opens up the game it makes it much, much more fun. Prior to the computers, top players could make arbitrary decisions about good and better positions and we could do no better than believe them. Now we know and we have found that sometimes the best players were, well, less than honest. But most of the time they were right.
The computers still have weaknesses. They still don’t understand one whit of strategy. It’s just that 90% of the moves in chess (and go) can be understood in concrete terms. It is funny to watch the engines get entirely lost and confused when a human player makes a very strong strategic move and the computer thinks it’s just a mistake, only to realize a few moves later that the human is better. If they have a good animation it can be fun to watch the engine go from “white is much better” to “black is winning” to “the position is equal” over and over as it gets lost.
That 10% of moves is not enough for a human to beat a computer, but it does highlight something that the algorithm doesn’t “get” yet. So it’s an interesting space.
Hawking has an agenda and he is smart enough to know what to say in order to get people focused on achieving his goals.
His goal is to get humanity focused on protecting our species from potential extinction. He is more worried about things like astronomical events, but if he can give us a short-term existential threat that the general populace can get worried about, then funds and manpower can be devoted to solving the (nonexistent) problem of AI while at the same time preparing for real threats.
In addition, it is a first step in thinking of ourselves as “species vs the universe” so we can be united for the cause, whatever danger we may face.
See Chapter 5, “An Enjoyable Game” : How HAL Plays Chess, from the book “HAL’s Legacy - 2001’s Computer as Dream and Reality” for a fascinating discussion of the state of chess computer programs prom the perspective of the books publication date, 1997. The chapter author, Murray S. Campbell, was a member of the original IBM team that developed Deep Blue, the machine that beat Garry Kasperov. In 1997, the prospect of a computer beating a Go champion was far, far off on the horizon.
At the end it states “Did you know… Go is the only board game in which humans can still defeat computers with reliable consistency”. It was dated 2010.
At this point, I will suggest reading Roger Zelazny’s greatest piece, a short story entitled “For a Breath I Tarry”. Foremost a tale of breathless beauty, its philosophical musings about AI and humanity have never been outdone.
I think that you are spot on with this post. Computers are jerks to humans. Also, I was surprised that a ~500mhz processor was about grand master level in chess. That makes AlphaGo more feasible in my mind, but also impressive. The number of GPU’s that it used, as you described it, was an insane amount. Humans seems really small when compared to how well computers can destroy us. The fact that AlphaGo lost also says something about humans not being able to create a perfect machine. Overall, I found this information really interesting.
These stunt matches are a bit unfair, because the computers have trained beforehand on the games of champions, but the human champion has never played against the computer. In both Kasparov and Lee, the human player was confounded by moves that no human had previously considered. Humans will train against computers, ushering in a golden age of human Go players.
AlphaGo seems to be better than Deep Mind. Deep Mind wasn’t really a man-vs-machine match: The engineers were tinkering with the program during the match, and promptly destroyed the machine at the end. Perhaps they were afraid of a rematch. AlphaGo is not available to the general public, but it’s built out of standard machine learning components. Perhaps patented, though.
we know that computers will continue to beat us at virtually every game we play
I think that’s true for a subset of games. I’m reluctant to think that a computer would win a game of Battlestar Galactica: The Board Game against my group.
For starters, he would be the very best to go to prison. He’s obviously a toaster and no ammount of good deeds would convince us otherwise… He could be in the winning team, granted, but he would be a burden. The team would win despite him, not thanks to him.
In conjunction with your newer post about coding games, and one of the comments on there - Advent of Code:
Many of the “part 2” problems on adventofcode (and some of the part 1) revolved around this change of how we program to solve problems. In other words, part 1 might be “find the best combination of weapons and armor for an RPG character.” And you could solve it with brute force. But part 2 would be “what’s the best order of spells to cast to survive a boss battle with the least amount of mana used?” And the combinations were so numerous that brute force was extremely likely to fail for most programming languages/machines, so you’d have to go at things a different way than you might expect. Maybe still a sort of brute force, but with randomization and aborting dead-end attempts to make things efficient.
Indeed. I didn’t spend much time identifying the type of problem because I’m not that sophisticated. Somehow I managed to solve them all with JavaScript anyway (occasionally leaning on help from fellow Redditors…) But my point was just that the two posts together reminded me of attacking those challenges, seeing how the brute force methods would fail, and then finding some way to let the program learn along the way that certain pathways were duds, so to keep trying something different. About 1/1,000,000th the level of learning that goes into neural network algorithms, but it just reminded me enough of it to comment about it
You cost per floating point operation chart is incorrect on the 1961 line. That should be $8,300,000,000,000 ($8.3 trilllion), not just $8,300,000,000 ($8.3 billion).
Jeff, not sure what you mean by “Pac Man is beyond the capabilities of Deep Mind”? Teaching a computer to play Pac Man is used as the end-of-semester test for college students at Berkley and is playable with some just some basic Q-Learning techniques: (https://youtu.be/jUoZg513cdE?t=40m54s). It’s also featured in the OpenAI playground, that anyone can code against.
I think you misunderstand. Deep Mind looked only at pixels, that is, it would have to figure out the game and its strategies from first principles based on the patterns of pixels on the screen.
If you give a computer a maze structure, and treat it like a maze, sure. That’s easy. But it is a far cry from saying “here’s a bunch of pixels, you figure out what they do.”
Details on the difficult games here: Montezuma’s Revenge, Pitfall, Solaris, and Skiing:
Here’s a neat summary of it learning to play breakout
Also, with regards to Fritz 9 and brute force benchmarks… 4 core, 17m, 8 core 28m, I just tested my fancy 16 core 5950x and got 38m. Diminishing returns with cores.
