As I write this, Lee Sedol lost the tournament, since he lost the first 3 matches out of 5. However, the format has Lee playing all 5 matches regardless, and it gets more and more interesting. There are 4 matches completed, and 15 minute summaries for the first 4. Here are the matches in order, and I will add more as it becomes available.
Enjoy! :)
Match 1, Pro Summary
Match 1, my dumb notes
Lee Sedol tried a very strange opening. It is not clear it has ever been seen in professional opening play. The theory is that he was trying to confuse the opening "book" database inside AlphaGo. AlphaGo doesn't work that way, so that's not a good reason. More importantly, black starts ahead. Black shouldn't risk losing his advantage; he should cement it.
Lee Sedol starts a corner fight early. He is playing a little too aggressively and he's a bit overconfident. However, starting fights early and attacking is very much Lee Sedol's style. It has worked against humans in the past. Again, only a 9-dan pro could even notice this stuff.
Lee Sedol sets up a very complicated fight, even by professional standards. In computer vs human game matches generally, it's not safe to increase complexity just to force a mistake. Some complexity, like "bookkeeping" a bunch of pending problems, can be done far better by our cyber overlords. When I beat a computer at Chess, I do it by playing a positional opening, and then forcing even exchanges to simplify the board. Then there is less chance that I make a mistake in the middle game. That said, this could work against AlphaGo. I call this the "blow the stack" approach. It is sometimes possible to increase the complexity to force the computer to run out of resources or time. You can tell when the stack is blown; the result is mistakes by the computer, sometimes looking like total nonsense!
Lee Sedol doubles down on the complexity. With more than four, simultaneous, raging, interacting battles in just one half of the board, this is a mind-boggling level of complexity. Alas, AlphaGo doesn't make any mistakes at this stage. This is very, very impressive, and never seen from a computer before! Michael feels guarded optimism with black's position in the middle game, but not "comfortable", due to the complexity.
White makes a move which is too small, sort of wasting a move. Later, black also makes a small move, which is somewhat wasted. Too many small moves allow your opponent to more rapidly deploy, claiming more territory than you. It's a race, where you play small moves to secure just enough of your existing territories, mixing in big moves to lay claim to larger tracts in the future. Of the two mistakes, black's mistake was worse. White relentlessly marched forward, taking a lot of territory from black as a result of this one mistake. Both white and black played excellently, by pro standards. A normal pro wouldn't stand a chance against either side in this game. AlphaGo wins this match. So much for Lee Sedol's boast that AlphaGo wouldn't win a single match!
Once AlphaGo "thinks" it's ahead, it will often play conservatively and even make sloppy moves or mistakes. In fact, sloppy moves often indicate that AlphaGo thinks it's going to win. The thing is, AlphaGo isn't trying to win as much as possible. It's trying to maximize the probability of winning by at least $\frac{1}{2}$ point. That's very different from humans vs humans. We often keep going, to test our opponents and to learn more about them. We also try to demoralize them. AlphaGo plays differently, in that sense.
Match 2, Pro Summary
Match 2, my dumb notes
AlphaGo leads a truly modern opening, and then quickly does an unusual move. Instead of shoring up a group of 3 black stones, black does a big move! It's almost like AlphaGo got another move for free. Sure, it makes securing the corner more difficult. But if AlphaGo can keep up the pressure, this will be very difficult for Lee Sedol as white. This is only 13 moves in!
Then, a few moves later, black reinforces the weak corner with a slightly unusual move, which is known to simplify the local situation later. Given that AlphaGo lost some initiative in that corner, by going big a bit earlier, this is actually a great way to go.
Play proceeds perfectly with both sides fighting in a nearby corner, and black keeping the pressure on throughout. Both players made big moves on the sides, which were both modern and normal. Then the bombshell!! Out of nowhere, on move 37, black attacks a single white stone with a single black stone on the side. This is a small move "too early" against a big move. Research indicates that there is only a 1 in 10,000 chance that a human would pick this move! In fact, Lee Sedol looked at the black stone for a minute, then left the room with the clock running...for 15 minutes! How's that for a surprise?
This link queues up the full-length coverage to the exact move where AlphaGo was brilliant. If you watch closely, Michael Redmond keeps moving the stone up and down in confusion. It turns out the move was typed incorrectly by the experts who enter the computer's moves into the record. Those experts, strong players in general, could't believe it either!
The rest of black's play is relentless. Michael couldn't find any mistakes in black's play, at all, and was surprised by several of black's moves. Black played smart, brilliant at times, and kept the pressure throughout. Black wins, leaving AlphaGo ahead 2-0. At the press conference, Lee Sedol, who sounds like a soft-spoken eunuch, remarked that he never felt in control during the entire match. Wow! Michael Redmond coined match 2, "AlphaGo's Exquisite Game".
Match 3, Pro Summary
Match 3, my dumb notes
Match 3 starts, with Lee Sedol playing black, and he chooses a modern, more "normal" opening. This is much better suited to his fighting style. Everything proceeds normally in the early game, until Lee Sedol chooses a very aggressive attack in the first corner. Michael Redmond believes that Lee Sedol could make this gambit work against most players in the world. Amazingly, AlphaGo will punish Lee Sedol for this later. That starts a couple dozen moves later. Time after time, white plays small moves in this corner which force black on the defensive, completely negating Lee Sedol's advantage as black.
Another commentator put the middle part of game 3 this way:
"After being compelled to flex its muscles for a short time and gaining the upper hand, AlphaGo began to play leisurely moves. By now, most observers know that this is a feature of the ruthlessly efficient algorithm which guides AlphaGo’s play. Unlike humans, AlphaGo doesn’t try to maximize its advantage. Its only concern is its probability of winning. The machine is content to win by half a point, as long as it is following the most certain path to success. So when AlphaGo plays a slack looking move, we may regard it as a mistake, but perhaps it is more accurately viewed as a declaration of victory?"
Michael notes that he could only find one small overreach on black's part. As he said; to have to unwind so many moves to find the source of the error is a testament to AlphaGo's strength.
AlphaGo wins, making the score 3-0. Lee Sedol has lost the tournament, and AlphaGo looks unbeatable. Or is it? All hail our Go AI rulers! Is there anything a human can do to win? By human, I mean Lee Sedol. It's clear nobody else would stand a chance.
This tournament has no "mercy rule", which means that the final 2 games will proceed as scheduled. Lee Sedol bowed and apologized, vowing to fight the last 2 games with all his might. Also, since Lee Sedol was too aggressive in the first few matches, it could be that AlphaGo surprised him. I wondered at this point, what would happen if Lee Sedol played more cautiously? What if Lee Sedol also tried to safely secure the minimal territory needed to win? It makes sense that a human might need a game or three to adapt. AlphaGo had the benefit of studying every human game, including all of Lee Sedol's. Prior to this tournament, Lee never saw a single game played by the new AlphaGo.
Can Lee Sedol adapt quickly enough to win even a single game? Is it game over for mankind?
We will see...
Match 4, Pro Summary
Match 4, my dumb notes
Lee Sedol plays white, so he moves second, at a disadvantage to AlphaGo. The game starts as a carbon-copy of match 2. Interesting. Maybe given the same moves, AlphaGo will tend to play the same responses? Very early on in game 4, white makes a break from the game 2 script. Instead of creating a neutral formation in the corner, white alters the placement of one stone, just a tiny bit. According to Michael Redmond, this puts more pressure on the trio of black stones, while also pushing up the right hand side. Then Lee Sedol sacrifices a single white stone, to gain a much more powerful left-side attack.
To me, this looks different already. Instead of attacking everything in sight, all at once, Lee is conceding stones and territory to gain some of his own. He's doing it on his terms, consolidating where he is strong first, before the game really even gets going. That's totally different. Will it work?
White plays a corner move which is a little small for Michael's taste. It seems again like too conservative a bid to get something solid in the corner. Then, along the side, white chooses to slowly and safely curl around and extend a small area. Black, in return, get's a thick wall aimed at the center. At this point I'm thinking, can Lee Sedol get AlphaGo to try to win the center? All things being equal, the center bid should fail. Maybe AlphaGo is so strong that it can do it? I'm secretly hoping AlphaGo tries for the center, or Lee Sedol pushes it around until AlphaGo "decides" to go for the center. If so, this is going to be the game I'd hoped for!
As Lee secured little islands around the edges, AlphaGo prepared to build a fortress in the center. Lee is preparing to later eat into all of black's positions. Lee is letting AlphaGo have some territory, but will try to eat into it, so it is not enough to win. As Lee is building nice shapes and formations, AlphaGo plays a bit too aggressively in the center. It's not an error. But it's not very defensible either. At this point, it looks like a large center formation will be created by AlphaGo, eventually.
White and black trade some more moves in the center. Suddenly at move 78, an amazing center move, wedging a white stone right in the middle of a black formation. This move took everyone by surprise. Michael exclaimed Wow!, when it happened live. Later, a Chinese 9-dan pro called it "a god move". Research also indicates that the odds of a human pro playing this move at 1 in 10,000. Michael believes that Lee Sedol must have had this move in the back of his mind, the entire time. So many of Lee's moves led up to it. This link to the full length video is queued up to the exact moment that Lee Sedol played his amazing move.
There was a problem though, Lee was running out of time. I think AlphaGo had about 45 mins left, when Lee ran out. When that happened, Lee has to basically move in less than 60 seconds. The center was a mess for black, but there was still lots of play left to go. Could Lee capitalize on the disarray he caused in the center? Would he run out of time? Would he make a mistake by moving to quickly? One thing was for sure, the pace was going to be quick now! How long can Lee play "speed Go"? It turns out, for at least an hour more.
The result was devastating. It looked like Lee "blew the stack" and AlphaGo had a partial lobotomy. There were a whole bunch of bad options in the center, and only one ok one. AlphaGo abandoned the entire troublesome area and played elsewhere. AlphaGo's challenges were all easily countered, and provided no distraction from the pending mess inside. At one point, Michael said that AlphaGo won't resign until it estimates less than a 20% chance of winning. He also said that if a pro were forced to play such "disheartening" moves, he would resign.
It went on and on. Black would get a point or two, but white would gain at least that much. If one bad move can lose a game for a 9-dan player, how could AlphaGo expect to survive a half-dozen, small, weak moves?
And then, Poof! AlphaGo resigned. Lee Sedol won, making the score 3-1 for AlphaGo.
The press conference was bedlam, by Go standards. Lee Sedol entered to thundering applause. Then, in his weak, girlish voice, Lee said he never received so much applause for winning a single game, after losing 3 straight. The AlphaGo team was congratulatory, but clearly baffled. Had Lee cracked the code? (pun intended)
Then, just at the end of the conference, another bombshell. Lee pointed out that since he had a win with white, it would be very valuable to try to win with black. Could he be permitted by the AlphaGo team to play black for the next game? The AlphaGo team agreed, to thunderous applause.
It's ON!!
Match 5, my dumb notes
This match was epic. It took 280 moves and 5 solid hours. It was nail-bitingly close, neck-and-neck throughout...except for move 48 by AlphaGo.
Here you can see the full-length video commentary, queued up to move 48 by AlphaGo. Michael Redmond was predicting this possibility, but then he notes that "it doesn't really work for white". It turns out that all pros study the formation facing the right edge of the board. It results in what pros call a "tombstone" shape, if played to conclusion. Michael Redmond and Lee Sedol can handle this case in their sleep. Amateurs will have seen the tombstone, yet maybe not remember exactly what it is. However, being able to see and identify this shape is one of the prerequisites for becoming a pro.
AlphaGo clearly made a mistake here, and it lost about 15 points in this area. I read one conjecture that this could have been part of AlphaGo's master plan (e.g. for the center). I disagree. AlphaGo, and computers in general, have problems thinking 20-30 moves ahead. The battle for the center is more than 100 moves from this move. Also, you can see AlphaGo "probing" the position, gradually "realizing" it's hosed. If AlphaGo recognized it's mistake or the resulting tradeoff, it would go attack another part of the board. There is then a chance that this could turn out differently later.
So, for the next 150 moves or so, Lee Sedol is considered to be leading a bit. First Lee Sedol runs out of time, and is forced to switch to "speed Go", with 60 seconds per move. Then AlphaGo runs out of time, triggering speed Go as well. The board is settled enough, so it's difficult, but Lee Sedol and AlphaGo don't make any apparent mistakes. The result is a tense, fast, war of inches. White grabs a space or two, black loses a space or two. Black grabs a space or two, and white loses as much.
Michael Redmond does not consider himself an amazing "counter". As a 9-dan pro, he is most certainly, fantastically quick and accurate at counting! More likely, his humility comes from the fact that the score is so close, and the board play is so complex. In the end, surprisingly, Lee Sedol resigns. He loses by just a few points; he almost won. The final score is 4-1 for AlphaGo.
Where did Lee go wrong? Nobody knows. The best theory I heard was from Michael Redmond, where it seemed Lee Sedol overreached, by one space! Slowly, it seems he just didn't grind out territory quite as efficiently as AlphaGo. At the press conference, Lee Sedol remarked, in a baby voice which would make Mike Tyson blush, that humans can beat the computers, but computers have superior stamina and concentration. I agree. We've seen this in the last two matches, once Lee Sedol learned to play more conservatively against AlphaGo.
Conclusion
That said, it plays big moves about as well as top human champions, which is amazing. Several 9-dan pros are looking forward to learning from AlphaGo, if and when it is released into the wild. One Korean pro remarked, that he felt stronger this week, after watching these matches. Surely, AlphaGo is capable of creating new, exciting, big moves. Hopefully, AlphaGo can train again, playing millions of games against itself. Then, perhaps using AlphaGo itself, cull those millions of games, looking for some new openings and also some new "god" moves. Pros studying these gems could herald a new era for Go. This is not idle speculation.
Powerful chess programs are used to train top chess players. Breakthroughs in Go knowledge happen regularly now, fostered by the Internet. Go professionals do play against computer servers, and against other players over the Internet. A public version of AlphaGo, running either inside Google or Amazon or Microsoft, could be the best Go study tool in history. It's just starting to get interesting.
In the end, I don't think the final score tells the story. Lee Sedol was turning the tide, adapting to AlphaGo. Lee Sedol is not used to playing conservatively, but he can. If AlphaGo remained "frozen" (no software updates), then it seems like Lee Sedol could someday beat it, perhaps half the time, maybe more. We don't know. We do know that human pro's can tell AlphaGo is not one of them.
There are times when AlphaGo play is indistinguishable from a human pro. Then there are times when it makes mistakes no pro would ever make. Of course, only the highest ranking pros can see this, like Michael Redmond and Lee Sedol. As long as we have pros of their caliber, the game isn't over yet.