Deep Blue v. Garry Kasparov
February 21, 2021 12:11 AM   Subscribe

The story of how a IBM exploited the passions and relationships of geniuses and then discarded them as soon as it stopped being profitable.

From the article: "IBM’s total control of the site and the playing conditions underscored the vulnerability of the human player. I was the only player in this competition influenced by any sort of negative or hostile atmosphere. I think IBM’s unwillingness to cooperate or give printouts of the computer’s thought processes harmed that atmosphere. (As of today, I still have not received the complete printouts that I requested.) There were also many minor incidents, starting with the fact that the venue was created for the convenience of the machine — with all these air-conditioning systems and dozens of people serving the machine — not the human player. I don’t want anybody to look at this as an excuse. It’s my fault. I accepted the conditions."

Here's a two hour video about Deep Blue. I highly recommend it - Fredrik Knudsen's videos are usually well-researched and well-performed.
posted by antihistameme (34 comments total) 20 users marked this as a favorite
 
There were also many minor incidents, starting with the fact that the venue was created for the convenience of the machine — with all these air-conditioning systems and dozens of people serving the machine

HVAC and staffing existing for a literal supercomputer was first of many "incidents"? The humans literally had furniture, air conditioning as well, and food and water, I assume. Maybe even staff to manage humans.

The man sounds like he found the experience traumatic because he lost.
posted by floam at 1:02 AM on February 21 [2 favorites]


The framing in the post is not present in the article, which is a contemporary account of the second match between Kasparov and Deep Blue. It's easy to read the article as trying to defend the integrity of the game of chess from the onslaught of computer solving; decades later, we know that the mere presence of artificial chess grandmasters was enough to chip away at the image of chess as a supreme intellectual pursuit.

Maybe the video does, but honestly the two hour bit puts me off.
posted by Merus at 1:10 AM on February 21 [1 favorite]


I have to say, I've never really understood the mindset of people who want to program computers to beat humans at games. It's like, wow, you managed to achieve computer dominance in an activity created purely for human entertainment. Nice job.
posted by nosewings at 1:23 AM on February 21 [18 favorites]


There may have been some kind of assumption in the chess community (I don't know, I am not part of it) that chess-playing computers provide all their calculations/logs in order to advance the sport both for humans and computers, as well as to show that the computer wasn't cheating. However, Deep Blue is mainly a marketing project to keep IBM in the news, not to improve chess or artificial intelligence. Any useful technology in the background would be patented and guarded and wouldn't be made available to others (without a substantial fee).
posted by meowzilla at 1:32 AM on February 21 [3 favorites]


I have to say, I've never really understood the mindset of people who want to program computers to beat humans at games. It's like, wow, you managed to achieve computer dominance in an activity created purely for human entertainment. Nice job.

I mean, you have to understand that when people started programming computers to play games, the computers were the underdog!
posted by atoxyl at 1:40 AM on February 21 [8 favorites]


There is a point to be made for your comment, floam - the next paragraph in the article reads: "This comment is not intended to doubt the integrity of the match officials, but rather to underline the concerns that might arise because of this. This condition did not, however, bother Kasparov in 1996 (when he won) and only showed up in 1997 (when he lost)."
posted by antihistameme at 1:41 AM on February 21 [2 favorites]


Any useful technology in the background would be patented and guarded and wouldn't be made available to others

To receive a patent, you must disclose the details of the invention and make that information used to receive the patent public. The protection is only in the commercial realm. That's the trade: share your info with everyone for a temporary monopoly, in the mean time anybody can try to build one in their garage if they want to.
posted by floam at 2:04 AM on February 21 [1 favorite]


I followed this closely at the time. It was intense. The first thing that happened is, Kasparov resigned after the computer played a great pawn sacrifice and then took ground him into a lost position. And then an amateur discovered that he could have gotten a draw, not long before he resigned! This was confirmed overnight by computers. The next morning, they had to tell Kasparov. The draw wasn't all that simple, but it is the kind of thing that the World Champion expects to see at once.

The match was practically over right there. First, that is a heavy, heavy thing to come back from in a short match - 6 games, iirc. But even more important was that he had missed the draw because he had already written off the game. The pawn sacrifice had spooked him. He thought it was an idea that a computer would not play. And he lost the last game that way too j he bluffed, allowing a sacrifice right in the opening that computers normally would be told not to play. Deep Blue crushed him with it.

Kasparov failed to overcome his misconception that he was playing essentially a big version of Fritz, the best PC program at the time. His paradigm for anti-computer training was play against Fritz. But Deep Blue was not the same kind of program. It was sort of an expert system for playing against Kasparov by then, with Grandmasters consulting in the tuning of its evaluations and planning its opening choices for the match, all implemented on parallel hardware that had custom move-generating chips and search hardware connecting it all. And it found lots of moves that the best PCs of the day needed many hours, or days, of computing to see.

This may have been what led him down the path of pretty much outright accusing IBM of cheating. He couldn't let go of the idea that only a human would play like Deep Blue had in certain positions. It was embarrassing. Not his finest moment.
posted by thelonius at 2:19 AM on February 21 [22 favorites]


It is an amazing series of human intellectual achievements to put together the knowledge to design and build machines that can apply mechanical sets of rules to thrash humans at board games. Mathematics, statistics, computer science -- not to mention the physical sciences & engineering to manufacture semiconductors. Even game-playing programs that identify novel game strategies that humans have never seen before are discovering and refining those strategies based on mechanical sets of rules that other humans have devised.

There's been an awful lot of progress in a small number of decades. Some of the foundational ideas of computing have only been around for less than 100 years. Shannon demonstrated a correspondence between boolean algebra and relay circuits in the 1930s.

Checkers: In the early 1950s Arthur Samuel at IBM developed a checkers program that was able to improve by playing against itself.

Chess (a lot harder than checkers): IBM's Deep Blue used a combination of minimax search with alpha-beta pruning + brute force parallel search implemented in special hardware + a heuristic evaluation function, tuned from analysis of games, to try to estimate the value of board positions.

Go (a lot harder than chess): DeepMind's AlphaGo combined monte carlo tree search (able to make tradeoffs between exploiting known high-value options with exploring lesser known options to gain information) with deep-neural-network based function approximation. AlphaGo was initially trained using a database of recorded historical games, and once it was sufficiently competent, trained against itself during self-play using reinforcement learning (in broad strokes this the same idea as Samuel's checkers program from the 50s). AlphaGo beat world class Go professional Lee Sedol 4-1 in 2016.

In 2019 Lee Sedol retired from professional Go play:

> “With the debut of AI in Go games, I’ve realized that I’m not at the top even if I become the number one through frantic efforts,” Lee told Yonhap. “Even if I become the number one, there is an entity that cannot be defeated.”

All that said, winning a board game like go is much easier than performing at human expert level in other less structured contexts such as driving cars.
posted by are-coral-made at 3:00 AM on February 21 [10 favorites]


realized that I’m not at the top even if I become the number one through frantic efforts

Sorry for the tangent, but this remind me of a growing feeling I have with digital computing and also the memetic culture of social media:

At the other far end of human experience is someone who enjoys toeing shapes in the sand by themselves because they like how it feels...but in between? What are we to make of human endeavor that is more than sensual play, but less than world renown (Sedol)?

I had a lot of competitive and non-competitive hobbies growing up and they have so much to do with identity. There was healthy amount of 'play' in them at all stages, but there was also achievement. How will the next generations of humans internalize this aspect of play when their peer group becomes not just the kids on their block, but everyone else on the planet who's online doing a given thing? How does one understand and develop exploration and improvisation when searching for existing results is clearly a faster track?

I know there are some good answers to these rather basic questions, but I also feel like we're becoming something else both at the individual and group level.
posted by Reasonably Everything Happens at 5:12 AM on February 21 [8 favorites]


How does one understand and develop exploration and improvisation when searching for existing results is clearly a faster track?

Capitalism really wrecks people from the inside out. Faster track to where?
posted by mhoye at 6:02 AM on February 21 [7 favorites]


Recently I heard Magnus Carlsen praised as the best player ever, as he plays moves most like what a computer would play.

Computers are good for learning as they can analyse your game and tell you where you went wrong.
posted by mokey at 6:40 AM on February 21


There's been an awful lot of progress in a small number of decades

If you look at the history of computer chess, there was really only a brief period where competition between top players and the strongest programs was meaningful. I think that the first program to play at about National Master strength was BELLE, about 1970. By the late 80's, Deep Thought, the precursor to Deep Blue (the project was renamed when IBM took it over, was able to play well enough that it could beat almost anyone, and matches pitting the computer against top grandmasters began to happen. And by the early 2000's, it was all over. The last couple of computer - GM challenges mostly featured the person exerting themselves to the utmost to try to hold the games to a draw. If they made any errors at all, they'd be mercilessly picked apart. So just about 12 years where these very interesting confrontations could really hold your interest.

Of course Go managed to compress this same progression into one match. I've never seen anything like it. There was no stage of Go software being almost good enough to win a match against an elite player but not quite, which was so exciting in chess.
posted by thelonius at 6:46 AM on February 21 [8 favorites]


There were also many minor incidents, starting with the fact that the venue was created for the convenience of the machine — with all these air-conditioning systems and dozens of people serving the machine

Don't human players spend a lot of time trying to 'psych out' each other during matches, finding things they can do (legally) to annoy the other person while they're figuring out their moves? Sounds like IBM found a way to distract Kasparov.
posted by Insert Clever Name Here at 8:37 AM on February 21


Now imagine IBM being able to record, copy and emulate interesting minds like the MMAcevedo short story posted earlier.
posted by loquacious at 8:49 AM on February 21 [1 favorite]


Now lets see how either Deep Blue or AlphaGo does against a 10 year old playing Go Fish. Both get a half hour to learn the rules.
posted by sammyo at 9:11 AM on February 21


Deep Calvinball
posted by scruss at 9:43 AM on February 21 [3 favorites]


Just to be clear, the linked article is from 1997 right?

I was never very impressed with Kasparov's whining that the Deep Blue tournaments were unfair. In any event the reality is within even a couple of years after he lost, no one thought any human would ever have a serious chance beating a computer with standard time controls ever again.

What I am impressed with is Kasparov's 2010s political campaign against Putin. He used his fame to stand up to a dangerous autocrat. He spent a few brave years working to undo Putin's grip on Russia only to finally be defeated and basically forced out of the country. So far he hasn't been poisoned.
posted by Nelson at 10:09 AM on February 21 [10 favorites]


Of course Go managed to compress this same progression into one match. I've never seen anything like it. There was no stage of Go software being almost good enough to win a match against an elite player but not quite, which was so exciting in chess.

The amazing thing about the Go engines compared to chess engines (at least, as I understand it) is that the chess engines play the sort of chess that humans recognize. They make a move and we say "That's an excellent move". The Go engines make moves that have experts scratching their heads and 23 plays later they see what's going on.

I vaguely remember a quote from someone who might be G H Hardy, but probably wasn't, about mathematicians. He said that some are like us, only better, and some (here he was probably talking about Ramanujan) exist on a different level. Alpha Go seems to be on a different level.
posted by It's Never Lurgi at 10:13 AM on February 21 [3 favorites]


How will the next generations of humans internalize this aspect of play when their peer group becomes not just the kids on their block, but everyone else on the planet who's online doing a given thing? How does one understand and develop exploration and improvisation when searching for existing results is clearly a faster track?

For the first question, I think YouTube already gives an answer to that. We're seeing kind-of-explosions of hobbies that I hadn't conceived of as a kid, largely enabled by online communities teaching each other, improvising, and raising everyone's skills as a result. The example closest to me is juggling, which has gone (over the last century) from a fairly close-guarded secret, with a relatively few people who knew the 'trick' of juggling a few balls, to a pretty giant online community trading tricks and tips and inventing all kinds of new things. Meanwhile, I have a neice who's super into decorative baking, which has a similar community around it. (Also, horror makeup - this video is from Everyday Cakes. These things overlap in interesting ways!)

For the second question, chess positions often have a single best answer, or at least a fairly constrained space of expression. As people and machines have gotten better at the game, the range of expression has grown smaller. It's hard to be a hypermodern if the machines know it's a losing strategy. Baking and music don't have this problem: it's an infinite space, with a collection of good moves that make interesting things possible, but ultimately the hobbyist gets to choose what they want to explore, without worrying about losing the game. Our games and hobbies don't have to have a weird zero-sum competition baked into them. And we get a lot out of communicating, learning, and mastering these things.

In short, i think the kids'll be alright.
posted by kaibutsu at 10:16 AM on February 21 [7 favorites]


I don't know if anyone has mentioned Kasparov's book Deep Thinking, which is a look at what he thinks the humanity-helping capabilities of AI are, woven into the context of the Deep Blue game. It's really worth a read in my opinion if you are curious about the match, as he goes into very great detail about the games and moves.

Two things that bear on this discussion are, first, that he said it's extremely tiring to play against computers, more so than people. They played even then so consistently at a high level that you couldn't think on your feet much, you had to devise huge 10-steps-ahead traps that took advantage of weaknesses in the computer's logic. He said exhaustion played a big part in his own blunders.

The second thing, and on preview Nelson has said he is not impressed with Kasparov's whining, but in the book as I recall he explains it pretty thoroughly. He says that he had found a flaw in Deep Blue's method that seemed fairly reliable, that it tended to do such and such and if you forced into a certain position it would choke, and I think he did that. But the IBM team, he claims, essentially saw what he'd done and patched out that vulnerability overnight! I can't remember how substantiated this is but they certainly performed maintenance and I believe admitted that they updated the logic.

So he does feel a bit robbed in that he felt he sort of played against two opponents: one he could beat, and another he couldn't.

But time has softened his feeling on the matter and he's come to feel pride in the fact that that was probably the last time in history any human player could beat an AI, and he was the last person who could, and did. He's very happily inverted the event's import so that he comes out rather on top instead instead of having the rather sad role he is usually cast into.

I'll add that I got to interview him on stage and he is a very, very sharp man. Every question I had he answered concisely and completely and without hesitation, so much so that I had burned through my cue cards with 90 seconds to go. I asked him my last question and he gave a brilliant answer that lasted exactly 90 seconds. As we went backstage after I explained how he had saved my bacon and he said though he didn't know I had no more questions, it did seem like a great time to stop and he saw the countdown timer and formed the perfect answer to end on. He's a keen one all right!
posted by BlackLeotardFront at 10:22 AM on February 21 [26 favorites]


...it's extremely tiring to play against computers, more so than people. They played even then so consistently at a high level that you couldn't think on your feet much, you had to devise huge 10-steps-ahead traps that took advantage of weaknesses in the computer's logic. He said exhaustion played a big part in his own blunders.

Another advantage that computers had in these matches is that they don't have any psychological state. The computer can take a terrible loss today and be back tomorrow good as new. The human, their confidence damaged and the pressure mounting, has to work through it and still try to perform at a level where they at least won't lose again.
posted by thelonius at 10:42 AM on February 21 [3 favorites]


> How will the next generations of humans internalize this aspect of play when their peer group becomes not just the kids on their block, but everyone else on the planet who's online doing a given thing?
I think that could go back to the time when people had to live in small groups of relatives—maybe fifty or a hundred people at the most. And evolution or God or whatever arranged things genetically to keep the little families going, to cheer them up, so that they could all have somebody to tell stories around the campfire at night, and somebody else to paint pictures on the walls of the caves, and somebody else who wasn't afraid of anything and so on.

That's what I think. And of course a scheme like that doesn't make sense anymore, because simply moderate giftedness has been made worthless by the printing press and radio and television and satellites and all that. A moderately gifted person who would have been a community treasure a thousand years ago has to give up, has to go into some other line of work, since modern communications put him or her into daily competition with nothing but the world's champions.

The entire planet can get along nicely now with maybe a dozen champion performers in each area of human giftedness. A moderately gifted person has to keep his or her gifts all bottled up until, in a manner of speaking, he or she gets drunk at a wedding and tapdances on the coffee table like Fred Astair or Ginger Rogers. We have a name for him or her. We call him or her an "exhibitionist."

How do we reward such an exhibitionist? We say to him or her the next morning, "Wow! Were you ever drunk last night!"
— Kurt Vonnegut, Bluebeard
(I don't entirely agree with this view but it sure has stuck with me...)
posted by aws17576 at 10:47 AM on February 21 [19 favorites]


Don't human players spend a lot of time trying to 'psych out' each other during matches

Definitely a characteristic of a certain kind of player.
posted by Rash at 11:39 AM on February 21


Ctrl+f "flying dildo" 0/0 :-(
posted by Token Meme at 12:50 PM on February 21 [2 favorites]


Baking and music are an infinite space to work in, but I do worry some about developing a social environment where we evaluate our creative work down to a better/worse scale, of audience engagement or algorithmic virality ranking.

The more our social interactions and discovery around a creative practice are mediated by an optimizing algorithm, the more I worry.
posted by away for regrooving at 5:14 PM on February 21


Computers are good for learning as they can analyze your game and tell you where you went wrong.

I would be curious to know if Magnus Carlsen-level players are using computers at a current or recent iteration of Deep Blue for discovering their own weaknesses and strengths.
posted by They sucked his brains out! at 10:54 PM on February 21


Constantly, but these days the program of choice is Stockfish.
posted by kaibutsu at 10:56 PM on February 21 [2 favorites]


The best computers nowadays have ratings in the 3500 range, whereas the best human players can do is the 2800s.

I saw an interesting interview with GM Mathew Sadler about a book he co-wrote called Game Changer, about DeepMind's AlphaZero chess engine, which has "uncovered an uncharted space of exploration within chess", and bested other engines with an attacking style unlike anything anyone's ever seen.
posted by mokey at 2:20 AM on February 22


Metafilter: Ctrl+f "flying dildo" 0/0 :-(
posted by Reverend John at 7:20 AM on February 22 [1 favorite]


One thing I noted at the time was that everyone was referring to Kasparov as the Reigning World Champion, when in fact, the official world champion at the time was Karpov. Even if you thought of FIDE as just a corrupt organization, it was the internationally recognized body in charge of the world chess championship. I understand why IBM and Kasparov wouldn't mention it, but I would have thought the media might mention it. You could say former world champion, or current champion of the (failing) organization he created- that seems like a pretty big asterisk. That's like forming your own Olympics and then declaring yourself an Olympic Champion.

I don't think IBM cheated, but there were certainly inherent advantages to Deep Blue. I don't know how to adjust fairness when it comes to the opening book. In standard human-human play, you're not allowed to bring opening notes, let alone the entire contents of ECO. So the human gets to use what's in his brain while the computer gets database access to millions of recorded games, along with evaluation. I believe Deep Blue also had access to the Nalimov tablebases for endgames.

And speaking of 'computer', he wasn't playing against one computer, but 480. (Or was it more? I thought it was over 1000). So Kasparov was playing against a team. I'm not sure he would have done much better if he had 479 helpers giving him notes throughout the game, but he probably would have drawn game 2.

There was also no way Kasparov could study Deep Blue's previous games, and even if IBM had provided test games, the programming could have been modified before the match. And the IBM team had access to all of Kasparov's tournament games.

More of a minor issue, but I don't see why humans were involved at all on the IBM side during the game. The machine should have been able to move the pieces itself, and been able to make its own calls about draws. (Although I think there was a game in the first match where the humans on the IBM team accepted a draw that the computer itself could have won)

/Whining. I don't think I wrote any of this down at the time, and it was looking for a way out of my brain.
posted by MtDewd at 8:44 AM on February 22 [1 favorite]


But time has softened his feeling on the matter and he's come to feel pride in the fact that that was probably the last time in history any human player could beat an AI

Someone should remind him that he wasn't the last world champion to lose to a computer, he was the very first. And that wasn't in 1996 vs Deep Blue - he lost to "Chess Genius" under rapid time controls in 1994. Two years later Deep Blue beat him under standard time controls, and that was the first time for that, too.
posted by floam at 12:26 PM on February 22 [1 favorite]


I would be curious to know if Magnus Carlsen-level players are using computers at a current or recent iteration of Deep Blue for discovering their own weaknesses and strengths.

Kasparov himself has been noted as one of the first high-level players to really take advantage of computers as a preparation tool, long before they could win a match against him.
posted by atoxyl at 1:40 PM on February 23


I think that the first program to play at about National Master strength was BELLE, about 1970.

1978 or so. Ken Thompson first wrote his chess-playing program for the PDP-11 in 1971, growing it and eventually adding dedicated hardware and naming it BELLE in 1976.

I can't find a citation for this, but it is quite likely that Ken Thompson was the first person in history to lose a game of chess to a computer.
posted by mhoye at 9:36 AM on February 25


« Older You Give Love a Bad Rasputin   |   Friendly Distance Bundle Newer »


You are not currently logged in. Log in or create a new account to post comments.