AlphaGo – The Movie [video]
youtube.com
youtube.com
In the Go world, the relationship between computer and human has turned into one of analysis. Every pro in Asia has a powerful GPU-equipped system with LeelaZero or KataGo or Minigo installed, and they use it to study openings and analyze their matches. LZ/MG are similar in that they're from the AlphaGoZero generation of bots, and thus are hardcoded to 7.5 komi and don't play well with handicap stones. KataGo is the open-source project with the newest ideas, using (among other things) territory ownership prediction, allowing it to handle arbitrary komi and handicap. Many pros are switching to it for this reason. All TV broadcast channels use AI for both winrate prediction and variation analysis. They'll report the predicted winrates by all major bots.
There are definitely AI-inspired opening sequences, but I would say that by and large, pro play has stabilized to just be better. A few years ago, pros would just mimic AlphaGo just to see what would happen, but there's becoming a better appreciation for the many different opening styles that are becoming possible, now that we have AI to analyze them. The strongest pros are the ones who spend the most time on their computers.
Despite a few extra years of research, all available AI systems still fail at ladders regularly. Some may have had ladder routines hardcoded in, but these routines still fail at loose ladders and the many wonderful ladder variants out there.
In the online Go world, cheating has become rampant. Top echelon of games/accounts are essentially just people competing on how beefy their GPUs are, and human professionals complain that they can't actually get a game against another person nowadays.
Professional tournaments / pro qualifying exams have had isolated incidents of cheating which are usually fairly obvious, as human playstyle is still recognizably different from computer playstyle.
haven't gotten objectively better from looking at it yet, but its super interesting. the ui is pretty nice, showing estimated probability vs move number - letting you find the places where the game swung one way or another any playing through those sequences.
I guess the only other thing I would ask for studying is move-by-move analysis for alternate sequences
1. The arrogance/confidence that people had towards AlphaGo assuming that it was not going to be competitive.
2. The lack of understanding tt even the programmers had for the game. At one point the devs were saying "I don't even know if that's a good move or not", yet their program is playing at superhuman play.
3. The amazement that commentators and other players experienced when AlphaGo performed move 37 in game 2.
4. The disappointment experienced by everyone when they came to the realization after game 3 the computer won.
5. The confusion when the computers neural network fails in game 4 and everyone (except the engineering team who sees the probabilities) is confused about whether the machines play is brilliant or terrible.
> 1. The arrogance/confidence that people had towards AlphaGo assuming that it was not going to be competitive.
This is pure stupidity. Those people didn't realize the nature of the game. Go is a (large) confined space, which, by nature, can be brute-forced for the best answer. ML was needed simply to trim and optimize search tree, to find the best local maxima from a broader range (than other AIs).
> 2. The lack of understanding tt even the programmers had for the game. At one point the devs were saying "I don't even know if that's a good move or not", yet their program is playing at superhuman play.
You can't even predict solutions from a simple DFS. Right things are not always obvious nor intuitive.
> 5. The confusion when the computers neural network fails in game 4 and everyone (except the engineering team who sees the probabilities) is confused about whether the machines play is brilliant or terrible.
It's likely that AlphaGo didn't look beyond "horizon"[1], and only realized it's losing after a dozen moves. In this sense, move 78 in game 4 itself is more amazing than move 37 in game 2. The only reason why the latter is uncommon is it's believed that it's difficult to build house in the sky(middle), thus one should keep stones close to ground(border), but that naturally depends on the situation.
Is there somewhere I could read a beginner-friendly explanation of what was so significant about the moves, and how and why they changed the games?
machines: I have won
humans: oh
AlphaGo showed the same dynamic: I don't know if everyone remembers, but before the Lee Sedol tournament, most 'sensible' 'respectable' people assumed that there was no way AlphaGo would win, any claims it would stomp Lee Sedol were hyperbolic AI alarmism, and the only question was whether DeepMind would be saved from acute embarrassment by eking out one victory over Lee Sedol or if he would simply clean the board with AG. And in the absolute worst-case scenario, it might be a fairly even match. After all, anyone could look at the Fan Hui vs AG games which were released a few months before and see that AG was really crappy, hardly even human pro level; Lee Sedol wouldn't break a sweat. Yes, it would improve if they trained it more and tweaked it more, but so what? There were so many ELO points between Fan Hui and Lee Sedol, and AIs lack human intuition. Just drawing a line is mindless curve-fitting, which sensible respectable people know better than to do.
Meanwhile, people like Eliezer Yudkowsky considered the training time and thought through the exponential graph, and concluded that, given all those months between Fan Hui and the tournament and the fact that DM was willing to do the tournament at all with so much money on the lines, AG would either lose terribly or could well go 5-0. As it happens, his prediction was wrong. It actually went 4-1, because Lee Sedol got lucky and accidentally triggered a 'delusion' (fixed in AlphaZero).
We went from 'the best MCTS might be somewhat competitive with a very low-ranked human pro head-on' to 'superhuman' in a matter of months, because AG performance scaled with compute. We've seen similar capability leaps with DoTA2 or SC2: the best non-cheating AI was so low level it basically didn't exist (DoTA2) or was nowhere near competitive with any pros (SC2) and then within a few months of finding a working architecture (compare Oriol Vinyal's demo at Blizzcon to the matches just a few months later with SC2 pros), they reach better-than-most-pro levels even if they are not strictly superhuman yet.
That's all fun and games when it's just board games or computer games, of course...
The surprise was that it was assumed that Neural Nets would scale exponentially like other computer projects - over the medium term, with a requirement to rewrite the program to take advantage of new hardware opportunities. The idea that the same program + 6 months could scale exponentially was pretty radical even to people who knew a lot about computers. I was certainly surprised because I'd made the assumption that AlphaGo could scale exponentially with improvements in graphics cards as opposed to exponentially with Google's commitment of computer hours. Obviously wrong in hindsight, but it wasn't a misunderstanding of exponentials, it was a misunderstanding of neural nets.
And we only really had one data point at the time, which was the Fan games. Couldn't even really draw a line without knowing a lot about the field.
In addition, real world domains may not be some parallizable.
So yes, current AI is like a brilliant intern. It will eventually be your manager, but its gonna take a few more years ?
Personally I'm not a fan of waxing, talking heads "documentaries" that tell us very little about what was actually done.
If you want to see alphazero games, there are a bunch of analysis of alphazero and stockfish games on youtube as well.
I think it's fair to say that a lot of "no-nonsense-just-give-me-the-facts" viewers in the HN demographic would find this type of film very unsatisfying. I.e. lots of interviews with virtually no technical explanations.
I quickly scrubbed the film and it looks like it's the same footage from the AlphaGo documentary I saw ~2 years ago:
https://www.imdb.com/title/tt6700846/
I don't know if this new video url dated "2020" is a different film or if they simply enhanced the 2017 documentary with new footage.
I disagree with this assertion. The film explores the idea that AlphaGo, like the game of Go itself, raises a lot of questions about what it means to be human. The relationship between AI and humans is developing and a topic of intense interest, especially among the HN crowd. I think what you might be saying is people with a high IQ but low EQ might not relate to this film. I'm sitting here wondering if people like that don't find human/AI intersection interesting because they are already mostly like a computer program themselves so it's not a big deal.
We're talking about 2 different things: the importance of the topic vs the the method of exposition.
You're emphasizing the topic's importance. I'm not disagreeing with that. Instead, I'm explaining that the documentary's style of exposition would be a turnoff to a subset of the HN demographic.
As example showing differences in how Battle of Midway can be explained...
Here's the "personality-driven" type of history documentary. Scrub the video and you'll see there's a lot of video showing the narrators walking around the beach and flying in helicopters:
https://www.youtube.com/watch?v=wo9wrUcb8jI
To be fair, some viewers need attractive people talking to the camera to make history "interesting". What does an attractive man walking along the beach have to do with explaining strategy and tactics at Midway?!? Nothing. Maybe some viewers need "eye candy" to go along with their medicine of history. Basically, all of The History Channel is structured as personality-driven type of documentaries with a human-interest angle.
Compare that to another style of delivery with no video of talking heads but has a voice-over with maps and higher information content:
https://www.youtube.com/watch?v=Bd8_vO5zrjo
If HN viewers want the 2nd style of documentary for AlphaGo, this thread's url is not it. The importance of "what it means to be human" is orthogonal to the method of explanation.
I wonder if people are looking for weaknesses or perhaps just ways to contribute, so that a human + AI could beat an AI playing alone.
Picture a brilliant but poor player. How would that person fare against someone who has a good AI to practice against?
Buying high end computer hardware is very cheap relative to the time required to become very good at go. Conversely, having a reasonably strong go program available on your mobile will mean there's more access to people to play and get analysis.
I would expect that AI makes it better for poor people. Previously, a non-poor person might take classes or have a tutor. Now, a poor person can have an AI analysis of each of their games if they want.
Indeed, the moves that a human would've overruled in the Lee Sedol tournament are the ones that vaulted AlphaGo to victory.
Lee is the real world's John Connor, the only man able to beat the machines.