Interesting how this is covered in the New York Times, while earlier, AlphaZero (and its predecessor AlphaGo) showed the similar kind of insight in playing go (a much more complicated game than chess) coming up with moves that humans would dismiss almost immediately. Since then, go playing professionals have started to imitate this style of playing. I guess the same will happen with professional chess players in the coming years: there will be a less strong focus on material and more on positioning. Because AlphaZero cares less about pieces, but more about their position and abilities to attach and/or defend.
IMO, at the highest levels of chess, there has always been a focus on "positioning" (positional chess) over material. World Champions like Capablanca, Botvinnik, Karpov, and Kramnik all play/ed in a style that was postionally sound, and at times boa-constrictor like. If you want to be a grandmaster today, you have to be able to understand/execute concepts like giving up material to establish a fortress ( https://en.wikipedia.org/wiki/Fortress_(chess) ).
The World Champion that played in the most sacrificial/attacking style, Mikhail Tal, was famed for giving up pieces to generate attacking momentum. Contemporary analyses of his play have found that some of these sacrifices were unsound, and some were actually the "best move" in a given position.
I don't think it's feasible to expect human players to be able to calculate at the ply/depth that AlphaZero (or other chess engines) is able to. See this example from the latest World Championship (https://www.chess.com/news/view/world-chess-championship-gam...). A "forced" win in 30 moves was available on the board, but it would've required that Caruana make moves that cut against the "principles" regarding piece placement ("positioning") that are drilled into chess players.
I think a simple reality is that the search depths that AlphaZero (and to a lesser extent other chess engines) are dealing with are simply beyond human capability. A human player trying to execute the sacrifices that AlphaZero did (https://chess24.com/en/read/news/alphazero-really-is-that-go...) would be taking a stab in the dark. In most positions, they wouldn't really be able to calculate all the variations, or foresee how the endgame would play out.
About the author (from the bottom of the article):
> Steven Strogatz is professor of mathematics at Cornell and author of the forthcoming “Infinite Powers: How Calculus Reveals the Secrets of the Universe,” from which this essay is adapted.
Is it just my sonar beeping off the charts, or does anyone else hear the unmistakable signs of a submarine article? (1)
Apologies in advance for what may be perceived as a rant. I have a very low tolerance for clickbait-y BS like this as it pertains to my own passions as a lifelong chess devotee and former professional player.
First, the author of the article has no professional credibility in either chess or machine learning. He's a professor of math and a writer. No disrespect to either math or writing, I love and value both very highly, but they have very little to do with chess and machine learning per se.
The problems is he tries to present AlphaZero as "humankind’s first glimpse of an awesome new kind of intelligence," which is really a bit of a stretch unless you add the disclaimer that technically all AlphaZero does is play 3 types of perfect-information games quite well. This is undoubtedly a great accomplishment, particularly in the field of Go which many domain experts felt intuitively would not crack to our AI overlords before another 5-10 years of computing power/hardware advances at least.
(As someone who had the unfortunate label of "prodigy" applied in my youth due to earning the title of chess master at age 10, I consider myself somewhat of a domain expert in chess, and I was one of those people who got it wrong. I barely know the rules of Go, but intuitively I could comprehend that it was several orders of magnitude more complex than chess, and I was really hoping that the Go gurus would fend off the machines for longer. They didn’t. Hats off to DeepMind.)
But. With all due respect to DeepMind engineers for an impressive result in chess and go, it's a bit too early to start thinking of AlphaXXX as an "oracle" where all we can do is "sit at its feet and listen intently" while we would "not understand why the oracle was always right" and eventually be left "gaping in wonder and confusion."
(As an aside, the amount of pseudo-religious worship language in the piece is truly off the charts. I realize it stokes the passions, but it would be great if we could talk about AI’s true strengths and limitations without resorting to such histrionics. But I digress.)
Why is it too early to start bowing down to a new god? Well, for starters, they basically just brute forced the game of Go a bunch of years earlier than predicted, but this wasn't just a pure software win, this was also heavily connected to massive increases in computing power aka GPUs and ginormous cloud-based render farms.
Secondly, the author tries to make the leap from AlphaZero [good at 3 perfect-information games: chess, go and shogi] to what he calls "a more general problem-solving algorithm; call it AlphaInfinity". Note how he invokes the holy grail of AGI (Artificial General Intelligence) without actually using this term, which would set off alarm bells in, well, anyone who knew anything about AI who wasn't employed by DeepMind/Google.
Notice further how this massive leap from "machine that can play 3 games well" to "machine that can, you know, actually think about stuff like a human can, including these pesky 'edge cases' and un-trained-for scenarios that always confuse our algorithms despite their otherwise inhuman level of perfection".
One great example of such a case that may cause one to question these glorious predictions is a research paper titled “Neural Networks Are Easily Fooled by Strange Poses of Familiar Objects” which shows how ML models consistently mistake a school-bus for a snowplow under the right (snowy) conditions (2). Far be it from me to dare bursting the bubble/reality distortion field of certain ML leaders and visionaries, but c’mon - a human child, once they truly learned how to recognize a schoolbus, would never mistake it for a snowplow, even if it was upside down.
This flaw doesn’t mean that we can’t update training data to handle these types of rotations, but it does mean that we have a lot of work to do before we can say that these ML models have in some way grasped the “essence” of “school-bus” or [insert-other-object] here in a deep symbolic way, and by "deep symbolic way" I mean "any way that a human child learns how to do reasonably quickly before moving on to other, exponentially harder tasks".
I could go on, but I won’t. Just in case my overall point isn’t clear:
1. AlphaZero is an unbelievably impressive accomplishment within the limited subset of life that is [chess, go, shogi]
2. ML approaches, even in computer vision, have a long way to go before anything remotely resembling child-level human intelligence
3. Therefore, can we please please stop the marketing masquerading as news articles about DeepMind’s latest result. And if anyone at DeepMind is listening: your product is pretty sweet! It would be better strategically to simply let it speak for itself, without trying to frame it as AGI.
> technically all AlphaZero does is play 3 types of perfect-information games quite well
You might be right about it being a submarine article but this seems to be underselling it. AlphaZero is (as I understand it) undisputed world champion by an indeterminate margin on the two most popular games, and has become so after only being given the basic game rules. If you told someone from 2015 that this would happen in the next decade, they'd laugh at you.
As I wrote near the top of my comment, I already laugh at myself for falsely predicting AlphaGo/AlphaZero dominance by a decade, give or take. But the leap from where we are now to AGI is several orders of magnitude more complicated, at least, than the leap from Chess to Go.
I think the issue here is that we started with "traditional" AI (logic solvers, inductive reasoning, graph search etc.) which were very clearly constrained and single-purpose. The goal has always been AGI which is all-singing-all-dancing post-Singularity magic. Now we have taken another couple of steps along the path from AI to AGI, we're gonna need another term for AI which can be applied to a range of tasks without significant further human intervention, but doesn't have the implied general reasoning and perception that our magic AGI will eventually have.
In spaceship terms, what we had was Soyuz and what we wanted was the Enterprise. Now we're about to get a BFR / BFS which can go to Mars and back. Is that a 'real' spaceship?
So apparently the +155, -6 score was against Stockfish 8. Stockfish 8 is rated by the CCRL list at 3379, with Stockfish 10 rated 85 points stronger than 8.
Worth noting that AlphaZero was only given 4 out of 9 hours of total training time when playing against Stockfish 8 (https://chess24.com/en/read/news/alphazero-really-is-that-go...), but I guess we can't make any real conclusions about AlphaZero vs Stockfish 10.
EDIT 2:
So apparently AlphaZero also "defeated" Stockfish 9, but the preprint of the upcoming paper in Science doesn't seem to provide a crosstable.
It seems that Stockfish 8 was given a 44-core machine to play on, and was not constrained in terms of time spent per move etc.
Generally an "undisputed champion" title would require a level of effort that the AlphaZero team apparently isn't interested in. They like to run matches between AlphaZero and Stockfish in private under their own controlled conditions, and publish the results. The rest of the computer chess world likes to run tournaments in full public view, and there are several entities which run tournaments between engines that way which can be used to gauge the relative strength of the engines.
What will be interesting is to see the effect on Leela (https://github.com/LeelaChessZero/lczero), which is public and open-source, takes part in public tournaments, and is built -- as much as possible, based on what's been disclosed in the papers -- to use the same approach as AlphaZero.
Yeah, fair point. I didn't even think to challenge him on that. Technically it's not undisputed given that DeepMind hasn't made it available for any kind of "fair fight" organized outside of its own office. I'm guessing AlphaZero would still win, assuming DeepMind isn't literally faking data, but the only way to be sure is a match between A0 and SF 10.
I've been pretty impressed with him as a sort of cross-disciplinary math wizard ever since I used his Nonlinear Dynamic and Chaos textbook in a class. His research is pretty wide ranging. He may be wrong about the future, but I'd be shocked if it was because of a lack of technical understanding of the algorithms.
You forgot to mention his sole grandmaster win came from a simultaneous exhibition where GM Larry Christiansen played against 35 people at once. Trust me, playing 30 boards at once is a wee bit of a handicap.
Not that I doubt his cross-disciplinary math skills or whatnot. But let's not pretend that he's anything resembling a serious chess player.
I don't think I forgot to mention that, I used a hyperlink to point to it instead. As a guy who loves chess and has studied it off and on over the years, I doubt I'll ever win a game against a grandmaster under any circumstances, so it still seems impressive to me.
But even I know enough, from watching videos on YouTube that analyze some of the AlphaZero vs. Stockfish games, to appreciate that AlphaZero was playing in a style I haven't seen before. What's the threshold for someone serious enough to write a chess article?
Saying "Strogatz did once beat a grandmaster in chess" without specifiying that it was simultaneous exhibition is a lie by omission. I am a chess player, and I would never insult another player by claiming to have beaten him or her without specifying the format. There is a way for you to say what you want to say without exaggerating to make a point.
What's the threshold for someone serious enough to write a chess article?
There's no threshold whatsoever, as long as you communicate a respect for the game and you don't try to use chess as a metaphor to make a larger point about AI and humanity that doesn't hold up to scrutiny. If a rank amateur can appreciate that AlphaZero plays chess beautifully and wants to write about it, by all means. I happen to agree. Just please don't abuse my beloved game for marketing purposes or to win an argument.
It's funny, articles like this might've inspired me a few years ago to go into this field, but the more that I learn about machine learning techniques, the more skeptical I get when reading things like this. But it's not so much the technical details that bother me but the ethical problem of hype. Hype itself isn't necessarily bad, but throwing it into public at best garners a bunch of harmless clicks and at worst is misleading.
I mean, I guess playing off of the whole "AI apocalypse" and the image of our computer overlords makes a quite evocative point. To be fair, this isn't just a problem with AlphaZero — similar problems in science journalism, especially in the field of medicine (i.e. Immortality is just a few decades away! Stem cells cure cancer!) seem to show up whenever a new or interesting result happens.
I thought "And when AlphaZero handicapped itself by giving Stockfish ten times more time to think, it still destroyed the brute." was a curious comment. Wouldn't it have been more accurate to say, "And when AlphaZero was handicapped..."?
Yeah. If you look closely, there are a lot of these little rhetorical hacks sprinkled throughout the article which has the effect of glorifying AlphaZero.
You are 100% right. For example, there is zero discussion of General Game Playing (GGP) in these submarine-type articles because the same techniques don't scale at all to GGP.
> First, the author of the article has no professional credibility in either chess or machine learning. He's a professor of math and a writer. No disrespect to either math or writing, I love and value both very highly, but they have very little to do with chess and machine learning per se.
Well, do you accept Garry Kasparov qualified to comment on this?
"I admit that I was pleased to see that AlphaZero had a dynamic, open style like my own."
"Alpha-Zero is surpassing us in a profound and useful way, a model that may be duplicated on any other task or field where virtual knowledge can be generated."
I know very little about AI, but something that excites me, maybe just romantically, is the idea of an AI taking everything it learned from one game and applying it to another.
Seems like we are happy to give them billions of games of practice. But what happens when exercise becomes a constraint?
You've played a ton of chess. Now here's the rules to Go. Now play one game of it.
The vast majority of "action" available online is "6-max" (6 player) or "full-ring" (9 player). On a PokerStars, WSOP.com, or Party Poker, you're going to find that there are maybe 1/10th or 1/20th the number of headsup tables as higher capacity tables.
The development of "GTO" (game theory optimal) play in Texas Hold 'Em is certainly a first step in the direction of computers playing poker. However, there's still quite a long way to go.
Here's the paper about Libratus (mentioned in the article).
In January 2017 Libratus beat a team of four top-10 heads-
up no-limit specialist professionals in a 120,000-hand Brains
vs. AI challenge match over 20 days.
The prospect of curing diseases is wonderful. But for now, we're just using these algorithms to figure out how to maximize the amount of time people spend staring at their phones.
I don't like the framing. AlphaZero is just a very refined and efficient form of brute force. Precomputing weights obtained by brute force doesn't make the overall enterprise not brute force anymore.
In your comment here and below it sounds like you're saying: "if you take a brute force algorithm and make it more efficient than brute force by not brute forcing the solution then it's just a brute force algorithm!"
If AlphaZero is brute force than any use of a non-exhaustive planning mechanism (pruned MCTS in this case) is brute force which is honestly ridiculous. Search and planning have a long history in both computer science _and_ neuropsychology because that is what we call the methods that are more efficient than brute force at the expense of some accuracy.
There are some problems with the article but it isn't that AlphaZero is just some overhyped brute force algorithm.
I never suggested it's overhyped. I think the specific terms being used to compare it to humans are inaccurate.
On a related note, AlphaZero used quite a lot of processing power - even with the optimizations, if it had to run on human hardware it would be pretty worthless.
It ran on a single machine with four TPUs. In a few years with a few more optimizations, I can imagine an equal strength implementation on a handheld chess computer.
If you're talking about training hardware, the correct comparison is against the processing time used by all serious human chess players through history because an apples to apples comparison would have both training from scratch (just the rules).
If we had a halting oracle and used it to output the best move for any situation, I would also call that brute force, even if it's actually only computing a handful of bits per move. Compressing a computation doesn't change the nature of what's being done meaningfully.
Your understanding of what brute force means is incorrect. It is an exhaustive algorithm. Any non-exhaustive algorithm is no longer brute force. An algorithm (like AlphaZero) that can locally generalize, even to small degrees, is also not brute force. Your argument is that a brute force algorithm is simply being compressed by a neural network but that isn't how AlphaZero works at all. You can't prune a brute force algorithm's branching and still call it brute force. In fact, if it was, then games like Go wouldn't be approachable by the same algorithm.
>One way to speed up a brute-force algorithm is to reduce the search space, that is, the set of candidate solutions, by using heuristics specific to the problem class.
The term brute force is clearly used for algorithms besides for ones that literally try every possible solution.
The term is clearly used widely to refer to more than just exhaustive algorithms.
The article portrays this as a battle between engines like stockfish, which are brute force (note that they explicitly call previous chess programs brute force, although clearly one based on your definition would be useless), against AlphaZero, which supposedly does more than brute force and has principles and insight etc. I think this gets it exactly backwards - stockfish had a lot of principles baked in by humans, as the article acknowledges, but alphazero was given 0 knowledge of chess other than the rules. Seems like this is properly viewed as a triumph of efficient algorithms over ones with principles, which is nearly opposite to the article's framing.
>>One way to speed up a brute-force algorithm is to reduce the search space, that is, the set of candidate solutions, by using heuristics specific to the problem class.
My reading of that sentence from Wikipedia is that this is the step by which a brute force approach is transformed into a "clever" approach. The hard part is to find valid and useful heuristics, which often also involve switching data structures in some way.
In what way is it brute force? Obviously it's physically infeasible to store the best move for every position. AlphaZero statistically learns a way to recognize the gist of positions in ways that were previously not possible; I see this as a great achievement.
It's a great achievement but I don't think framing it as thinking akin to humans is a valid framing.
>AlphaZero gives every appearance of having discovered some important principles about chess, but it can’t share that understanding with us.
No, it doesn't have any new principles, it just has a very good system of weights.
Imagine a perfect computer that plays the game by mapping out every single move. AlphaZero is a series of optimizations that allow for an efficient but lossy simulation of such a computer. Most of the research in this area is about making an oracle more practical/more accurate estimations of the theoretically correct response. I think it's fair to call it brute force with optimizations.
You could say the same about the way humans have learned to play chess. The main lines of the popular openings have been exhaustively searched. The way humans evaluate the middle game is also due to distilling features from brute force play and passing them on to other players who add on their own features and compute weights from additional brute force play.
The difference between AlphaZero and StockFish is that StockFish only does the latter (compute weights for features given to it by others), while AlphaZero also does the former (distill features from game state).
A good analogy is intuition. If someone says AlphaZero has developed a strong intuition for chess, I have no issue with that, and to the extent human chess is based on intuition comparison is fair. But to portray humans as having chess principles, then say AlphaZero has new ones but just can't describe them seems off.
To put it another way, humans have general intelligence and AlphaZero doesn't. At least some of the general intelligence is used for human chess play, so there is a dimension of that which AZ doesn't have.
In the context of playing chess, humans use general intelligence to featurize the game state in terms of piece value and position value and weight the relative importance of those. The rest is memorization and simulation, which is where the intuition of advanced players comes from — they aren't actively thinking of piece value and position rules like beginmers. I don't see a huge difference here between humans and AlphaZero in those terms.
If you look at the progress graph of AlphaZero / LeelaChess (open source clone), and look at how it plays in its weaker iterations, it actually pretty closely correlates to how a weaker human chess player would play at those levels of strength (positional blunders, simple tactical mistakes, etc).
I've played chess for many years and I've played a few games against Leela (not full strength, the early versions). I never got the feeling that I was playing a computer.
Additionally, one of the primary ways of improvement past the ~2000 level is review of master games. Most people who make it to GM level have reviewed upwards of 5,000 master games or more over a significant period of time. So they build up a similar system of "weights" in their mind for different positional features, very similarly to the way the NN would approach the problem.
Sure, AlphaZero/Leela has not obtained generalized intelligence, but within a constrained sphere (Chess, Go, etc), it has come just about as close as is possible to that.
Reviewing 5k games is nowhere near the level of computation required for AlphaZero. They would likely not even be able to beat a human if running on human level hardware. I think there's a long way to go, and a future, general algorithm would be at AlphaZero while running on a laptop
Probably. its almost like a "poor man's hash" where you can retrieve the "almost similar best response" for each move. It stores all the games it has self-played in it's network and then looks up best move based on that
