> I am not sure I believe this explanation! When I look at subreddits for Go players who use Lizzie, my impression is that they don’t look at the reasoning all that much. They use it mainly to pinpoint moves where the winrate suddenly drops, so they can focus their learning on their biggest mistakes.
I think you should discount those observations a bit. The way typical players (mostly beginners and casual players) on reddit are using AI analysis in Go will not be representative of top players. I'm mid-amateur dan and still far from the top, but closer enough to stronger players that I can perceive myself some of that from my own personal experience. If you want to get a better impression of how strong players think about AI analysis, take a look at Michael Redmond's streams (9 dan pro) where he analyzes various games of both his own, or AlphaGo's games, with mention about various AI-suggested alternatives - it's not just looking for drops and parroting moves, but rather often diving deep into variations to place it into the context of his experience with similar positions.
> It is Shin, Kim and Kim who claim Leela Zero helped because, unlike AlphaGo, it showed the reasoning behind the move, not just the move.
> The true explanation why open source helped might actually be the inverse of what Shin, Kim and Kim propose. It might that the reason open source helped was that it let people do massive input learning
I don't recall who Shin, Kim, and Kim are, but assuming they're on-the-ground-informed about how players use AI in the same kinds of ways I've observed myself, then it's possible you might be misinterpreting what they are saying in a way that makes it more opposed to your proposed "true explanation" than it really is. There's a different interpretation that is not contradictory to your hypothesis. Which is that:
* Seeing just the isolated move that a strong AI proposes in a given situation is not so useful for learning. It's extremely hard to guess what situations that move generalizes to or not - slight changes to the surroundings can easily change the best moves.
* But seeing the all the sequences of moves that a strong AI proposes including all the relevant counterfactual sequences, is more useful for learning. e.g. "The AI proposes X, but the opponent can just respond Y, that seems bad for me? But the AI doesn't have the opponent respond with Y, it concedes and trades with Z! So presumably it thinks Y is not a refutation. Let me force X-Y and analyze again from there... aaah I now I see that Y fails because such and such stone is present. Now my brain is trained with the exact stone/shape/tactic to look for that makes X possible." And a dozen other different flavors of different kinds of counterfactuals that you could ask.
The latter is only possible if you actually can scroll back and forth through variations and interrogate the bot on different sequences interactively in different situations, which is only possible with e.g. a Leela Zero, and not just a static set of AlphaGo game records. And my own experience is that it actually is a big help, so long as you are independently strong enough at the game to be capable judging enough of the answers you get back when interrogating different sequences.
If you interpret Shin, Kim and Kim's "the reasoning behind the move" as referring to seeing the full sequences and counterfactual sequences, and not as referring to the low-level mechanism of learning - then there is no conflict with your hypothesis. Seeing counterfactual sequences and refutations and interrogating the bot interactively where you were unsure can be a big help for learning at the *same time* as the mechanism of that learning could be mostly pattern recognition training through lots of data. Indeed, seeing all those sequences is part of getting that concentrated data in order to train one's pattern recognition!
> After a few years, the weakest professional players were better than the strongest players before AI. The strongest players pushed beyond what had been thought possible.
I think you are misinterpreting this graph, looking at the SSRN paper (https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3893835). They do not make any statements I see about the population distribution Elo before/after or that the improvement means that an ordinary professional could now beat Ke Jie or Lee Sedol in their prime. This graph seems to just be about the average move quality across the whole Go player population increasing a bit per move. The lines are not the population, but simply the uncertainty around the mean. (This would be like estimating the American population at an average of 5-foot-7 with a standard error of 0.1 inches, and concluding that basketball players are impossible; or that if you measure the Dutch population at 5-foot-9 +- 0.1 inch, every single Dutch person is taller than every single American person; or that after a bunch of health interventions during the 1900s, the American population mean increased by 1 inch and then all young Americans were taller than all old Americans.)
My brother is 13 and played over 2000+ (online) chess games last year. He spends at least 2-3 hours a day on chess.com and I think that's a conservative estimate... When I play with him I sometimes ask him to think out loud and he basically does a chess.com play-by-play analysis of our moves.
Not on the content, but on the form: I think it would have been better to move part of Note 2 into the main text. If I only read the text, I would totally agree with the whole reasoning thesis of SKK. It was only by reading Note 2 that I got to the counter-intuitive and beautiful idea that, actually, the "just do this way" might unlock new levels of development. This seems to me to be important regarding how AI might help us.
EDIT: Just read polytope's comment and I would also a summary of his thoughts to the main body of the text, crediting him, of course.
This is fascinating. I wonder what's going on here. Is it seeing 'the art of the possible' that raises motivated expectation, or is the study of how AI plays Go creating greater sophistication and variety in the way Go players are creating their mental representations of the game? Or is there something spooky at play as in parallel innovation when two discoveries take place simultaneously when there is no discernible way for the information to be transmitted other than through explanations relying on a collective consciousness or a Sheldrake's idea of Morphic Resonance.
I tried to listen to a Boulez piano sonata while reading the article - I have fairly eclectic tastes but found the extraneous cognitive load of the auditory interruptions a little overwhelming.
I’ve been writing my whole life. I write professionally. I write in my free time for fun.
I have never been better at writing because of ChatGPT.
It isn’t about slapping together a prompt and then copy/pasting it. ChatGPT can write the bad first draft. Give you ideas for the outline. Tell you what isn’t working too well. Help you brainstorm. Become your personal editor. If used properly, it can be your writing mentor.
I have a hunch AI writing tools will be a staple of professionals at any level within our lifetime.
Me too! I don't find it useful for outlining or writing/style. But I use it to pinpoint holes in my logic and tell me if I miss some angle. I also use it as a fuzzy synonym search, which is useful since my english vocabulary is rather more limited than my Swedish so I can think a lot more things than I can say but can't google them. Also, voice transcription has been a big step up recently since I've started using whisper locally to sketch out parts of first drafts somtimes + record feedback sessions with my editor + random conversations to better remember ideas etc. And I also use chatgpt to summarize those things so it is easier for me to remember what was in a particular converastion.
I suspect this will add up to me being able to push my craft further faster, focusing on the elements I add - insight, personality, taste, personal experience, weird input data that is out of distribution etc.
You see, the high jump is a simple event. The athletes jump over a bar and whoever jumps the highest wins the event. Usually, each athlete will toss their body over the bar and crash onto a padded landing pit on the other side. Like most schools in the 1960s, the landing pit at Fosbury’s high school was made of wood chips and sawdust. Before his junior year, however, Fosbury’s high school became one of the first to install a foam landing pit and that gave him a crazy idea.
What if, instead of jumping the conventional way with his face toward the bar, Dick Fosbury turned his body, arched his back, and went over the bar backwards while landing on his neck and shoulders?
Dick Fosbury found success because his sport had switched the landing material and he was willing to experiment with a new jumping style. Let’s consider some common situations where experimenting with new approaches would serve us well.
"My guess is that AlphaGo’s success and alien playing style forced the humans to reconceptualize the game and abandon weak heuristics. This let them see possibilities that had been missed before."
My own guess is that the changes are more subtle. In certain situations, giving a slightly higher weight to the factors that favor move X over move Y than you used to.
They run 10,000 simulated games starting from that move to see how it affects average win rates. There is a limitation in this measure in that it to some extent measures how much like the AI they play, but since this way of playing also lets them win more it is not a goodheart law situation (even if it might be a few percent of that).
Interesting! I'd be curious to see what happens more years down the line, if/when AI caught up with the players again what would that do
On another note, in running a training at work recently i have been really feeling that sense of shifting people's windows of possibility through challenge and example, and this article reminds me of that
I didn't think the players had caught up with the AI yet - in Go and in Chess, people are playing better than ever but we'll never again see a human who can beat the computer.
Correct. But in this domain, as in a bunch of others, it doesn't matter isn't that people aren't best; it is that their play is more interesting and rich. No one cares to see a computer play. Interesting to see how many domains have this property.
How sure are we that this is a real increase? The worst players now being better than the best players were in 2016 seems extreme; surely some of them weren't using AI to train?
These are all professional players! So it seems reasonable that the someone ranked 100 can climb to the level of someone ranked no 1 after five years of better training. It would be interesting to see if anyone refuses to use AI - no idea.
Hello, Henrik, thank you for the interesting article! I do not play Go, but have some background in chess, and read it from the perspective of a chess player. I am curious about your statement that there are some chess players who recent became grandmasters who did it "largely by playing 10+ hours a day of online speed chess instead of the older strategies that emphasized targeting learning, deliberate practice and slower exercises". I was unable to find interviews with such players supporting the point. In fact, every single titled chess player I've heard/read so far, who has been asked how to improve in chess, answered exactly the opposite, that one should prioritize playing slow time controls to improve. Can you please provide your sources those lead you to writing the statement? Thank you.
I recalled a similar thing with Poker as it experienced a moneyball-esque revolution in the early 2000s W online play, Data and Game theory. It opened up barriers to the game (internet made info accessible to all) but raised the quality of the game even higher
> I am not sure I believe this explanation! When I look at subreddits for Go players who use Lizzie, my impression is that they don’t look at the reasoning all that much. They use it mainly to pinpoint moves where the winrate suddenly drops, so they can focus their learning on their biggest mistakes.
I think you should discount those observations a bit. The way typical players (mostly beginners and casual players) on reddit are using AI analysis in Go will not be representative of top players. I'm mid-amateur dan and still far from the top, but closer enough to stronger players that I can perceive myself some of that from my own personal experience. If you want to get a better impression of how strong players think about AI analysis, take a look at Michael Redmond's streams (9 dan pro) where he analyzes various games of both his own, or AlphaGo's games, with mention about various AI-suggested alternatives - it's not just looking for drops and parroting moves, but rather often diving deep into variations to place it into the context of his experience with similar positions.
> It is Shin, Kim and Kim who claim Leela Zero helped because, unlike AlphaGo, it showed the reasoning behind the move, not just the move.
> The true explanation why open source helped might actually be the inverse of what Shin, Kim and Kim propose. It might that the reason open source helped was that it let people do massive input learning
I don't recall who Shin, Kim, and Kim are, but assuming they're on-the-ground-informed about how players use AI in the same kinds of ways I've observed myself, then it's possible you might be misinterpreting what they are saying in a way that makes it more opposed to your proposed "true explanation" than it really is. There's a different interpretation that is not contradictory to your hypothesis. Which is that:
* Seeing just the isolated move that a strong AI proposes in a given situation is not so useful for learning. It's extremely hard to guess what situations that move generalizes to or not - slight changes to the surroundings can easily change the best moves.
* But seeing the all the sequences of moves that a strong AI proposes including all the relevant counterfactual sequences, is more useful for learning. e.g. "The AI proposes X, but the opponent can just respond Y, that seems bad for me? But the AI doesn't have the opponent respond with Y, it concedes and trades with Z! So presumably it thinks Y is not a refutation. Let me force X-Y and analyze again from there... aaah I now I see that Y fails because such and such stone is present. Now my brain is trained with the exact stone/shape/tactic to look for that makes X possible." And a dozen other different flavors of different kinds of counterfactuals that you could ask.
The latter is only possible if you actually can scroll back and forth through variations and interrogate the bot on different sequences interactively in different situations, which is only possible with e.g. a Leela Zero, and not just a static set of AlphaGo game records. And my own experience is that it actually is a big help, so long as you are independently strong enough at the game to be capable judging enough of the answers you get back when interrogating different sequences.
If you interpret Shin, Kim and Kim's "the reasoning behind the move" as referring to seeing the full sequences and counterfactual sequences, and not as referring to the low-level mechanism of learning - then there is no conflict with your hypothesis. Seeing counterfactual sequences and refutations and interrogating the bot interactively where you were unsure can be a big help for learning at the *same time* as the mechanism of that learning could be mostly pattern recognition training through lots of data. Indeed, seeing all those sequences is part of getting that concentrated data in order to train one's pattern recognition!
Oh! Thanks that makes things a lot clearer to me. I'll pin this comment.
> After a few years, the weakest professional players were better than the strongest players before AI. The strongest players pushed beyond what had been thought possible.
I think you are misinterpreting this graph, looking at the SSRN paper (https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3893835). They do not make any statements I see about the population distribution Elo before/after or that the improvement means that an ordinary professional could now beat Ke Jie or Lee Sedol in their prime. This graph seems to just be about the average move quality across the whole Go player population increasing a bit per move. The lines are not the population, but simply the uncertainty around the mean. (This would be like estimating the American population at an average of 5-foot-7 with a standard error of 0.1 inches, and concluding that basketball players are impossible; or that if you measure the Dutch population at 5-foot-9 +- 0.1 inch, every single Dutch person is taller than every single American person; or that after a bunch of health interventions during the 1900s, the American population mean increased by 1 inch and then all young Americans were taller than all old Americans.)
updated
Thank you! I think you are right. Pinning this and will edit if I find the time.
My brother is 13 and played over 2000+ (online) chess games last year. He spends at least 2-3 hours a day on chess.com and I think that's a conservative estimate... When I play with him I sometimes ask him to think out loud and he basically does a chess.com play-by-play analysis of our moves.
Not on the content, but on the form: I think it would have been better to move part of Note 2 into the main text. If I only read the text, I would totally agree with the whole reasoning thesis of SKK. It was only by reading Note 2 that I got to the counter-intuitive and beautiful idea that, actually, the "just do this way" might unlock new levels of development. This seems to me to be important regarding how AI might help us.
EDIT: Just read polytope's comment and I would also a summary of his thoughts to the main body of the text, crediting him, of course.
This is fascinating. I wonder what's going on here. Is it seeing 'the art of the possible' that raises motivated expectation, or is the study of how AI plays Go creating greater sophistication and variety in the way Go players are creating their mental representations of the game? Or is there something spooky at play as in parallel innovation when two discoveries take place simultaneously when there is no discernible way for the information to be transmitted other than through explanations relying on a collective consciousness or a Sheldrake's idea of Morphic Resonance.
I tried to listen to a Boulez piano sonata while reading the article - I have fairly eclectic tastes but found the extraneous cognitive load of the auditory interruptions a little overwhelming.
lol to the Boulez soundtrack.
I think it is mainly studying the AI and getting deeper less heuristic based mental models of the game.
Morphic Resonance
I’ve been writing my whole life. I write professionally. I write in my free time for fun.
I have never been better at writing because of ChatGPT.
It isn’t about slapping together a prompt and then copy/pasting it. ChatGPT can write the bad first draft. Give you ideas for the outline. Tell you what isn’t working too well. Help you brainstorm. Become your personal editor. If used properly, it can be your writing mentor.
I have a hunch AI writing tools will be a staple of professionals at any level within our lifetime.
Me too! I don't find it useful for outlining or writing/style. But I use it to pinpoint holes in my logic and tell me if I miss some angle. I also use it as a fuzzy synonym search, which is useful since my english vocabulary is rather more limited than my Swedish so I can think a lot more things than I can say but can't google them. Also, voice transcription has been a big step up recently since I've started using whisper locally to sketch out parts of first drafts somtimes + record feedback sessions with my editor + random conversations to better remember ideas etc. And I also use chatgpt to summarize those things so it is easier for me to remember what was in a particular converastion.
I suspect this will add up to me being able to push my craft further faster, focusing on the elements I add - insight, personality, taste, personal experience, weird input data that is out of distribution etc.
I think that’s the most empowering part about all these tools. Anyone can use them the way that best suits their ability.
"Now [Boulez] is standard repertoire at concert houses." Which ones?
I gather that piano sonata no 2 is most popular. Not sure which one is considered most unplayable.
https://www.youtube.com/watch?v=bnB94tZFwZE
Another similar examples come from
High jump:
https://jamesclear.com/dick-fosbury
You see, the high jump is a simple event. The athletes jump over a bar and whoever jumps the highest wins the event. Usually, each athlete will toss their body over the bar and crash onto a padded landing pit on the other side. Like most schools in the 1960s, the landing pit at Fosbury’s high school was made of wood chips and sawdust. Before his junior year, however, Fosbury’s high school became one of the first to install a foam landing pit and that gave him a crazy idea.
What if, instead of jumping the conventional way with his face toward the bar, Dick Fosbury turned his body, arched his back, and went over the bar backwards while landing on his neck and shoulders?
Dick Fosbury found success because his sport had switched the landing material and he was willing to experiment with a new jumping style. Let’s consider some common situations where experimenting with new approaches would serve us well.
"My guess is that AlphaGo’s success and alien playing style forced the humans to reconceptualize the game and abandon weak heuristics. This let them see possibilities that had been missed before."
My own guess is that the changes are more subtle. In certain situations, giving a slightly higher weight to the factors that favor move X over move Y than you used to.
That makes sense.
Good article based on assurance and encouragement rather than fear and trepidation. More, please!
Great article. Can you tell me - in the graphs posted, how was decision quality measured? Thanks!
They run 10,000 simulated games starting from that move to see how it affects average win rates. There is a limitation in this measure in that it to some extent measures how much like the AI they play, but since this way of playing also lets them win more it is not a goodheart law situation (even if it might be a few percent of that).
Interesting! I'd be curious to see what happens more years down the line, if/when AI caught up with the players again what would that do
On another note, in running a training at work recently i have been really feeling that sense of shifting people's windows of possibility through challenge and example, and this article reminds me of that
I didn't think the players had caught up with the AI yet - in Go and in Chess, people are playing better than ever but we'll never again see a human who can beat the computer.
Correct. But in this domain, as in a bunch of others, it doesn't matter isn't that people aren't best; it is that their play is more interesting and rich. No one cares to see a computer play. Interesting to see how many domains have this property.
This isn't true. Humans can currently defeat top go programs by exploiting a blind spot they have. See https://goattack.far.ai/
Ah! Thanks for the clarification
How sure are we that this is a real increase? The worst players now being better than the best players were in 2016 seems extreme; surely some of them weren't using AI to train?
These are all professional players! So it seems reasonable that the someone ranked 100 can climb to the level of someone ranked no 1 after five years of better training. It would be interesting to see if anyone refuses to use AI - no idea.
This isn't true. Humans can currently defeat top go programs by exploiting a blind spot they have. See https://goattack.far.ai/
Hello, Henrik, thank you for the interesting article! I do not play Go, but have some background in chess, and read it from the perspective of a chess player. I am curious about your statement that there are some chess players who recent became grandmasters who did it "largely by playing 10+ hours a day of online speed chess instead of the older strategies that emphasized targeting learning, deliberate practice and slower exercises". I was unable to find interviews with such players supporting the point. In fact, every single titled chess player I've heard/read so far, who has been asked how to improve in chess, answered exactly the opposite, that one should prioritize playing slow time controls to improve. Can you please provide your sources those lead you to writing the statement? Thank you.
nihal, narayanan, etc
I heard it from Nabeel Quereshi whom is much more in the know about chess than I am
I recalled a similar thing with Poker as it experienced a moneyball-esque revolution in the early 2000s W online play, Data and Game theory. It opened up barriers to the game (internet made info accessible to all) but raised the quality of the game even higher