Posted by gmays 1 day ago
In go ELO like scoring he’s something like 120 points over the next strongest player. No other player has ever broken a 3800 rating let alone 3850. Ke Jie (the previous long time champion) peaked at 3755. Shin Jinseo’s strength graph is the most absurd straight line.
2 stones is historically the gap between a 9P ranked and a 1P ranked professional player (very roughly the gap between super grandmasters and an almost grandmaster)
That is to say it’s shocking that Katago (almost certainly significantly stronger than AlphaGo) is a mere 2 stones stronger than Shin Jinseo. I suspect it would be 3-4 stones vs any other human pro.
And of course you would need to take into consideration the scale of go ratings and chess ratings when making that comparison. With top chess ratings being around 2800, being 1000 less than the top go ratings, one would have to apply a factor of roughly 3/4.
It's not a comparison of the worth of the games (I play both, though I'm better at Go, and prefer it), but the dynamic range of Go is larger.
That said, any cross-game/sport comparisons of this kind are pretty tough to do properly.
English is not my first language but for clarity perhaps the title should be:
Go grandmaster Shin with a two-stone handicap defeats AI KataGo
However, you can't compare goratings over time, the top ranks are not nearly stable enough. https://www.goratings.org/en/history/ (I think it's believable Shin Jinseo is better than Lee Changho, but not that there has been steady progress since the days of Lee Changho, so that there are now 20 players stronger than him).
Not without flaws of course, but probably interesting
Interesting KataGo is an open sourced Go program written primarily by David Wu in C++ and recently heavily vibe coded by Claude. It's running on four Nvidia RTX-3090 GPUs with 96GB VRAM. [1]
Based on the youtube video, it looks like katago was only using 16 seconds per move, is that right? https://www.youtube.com/watch?v=-86zF4mTWOY
Is 20 seconds on that hardware really overkill and well into the diminishing-returns curve, as a top-level comment suggested, or is it plausible katago could have played better if given 40 seconds per move?
match details: https://gostonebase.com/blog/shin-jinseo-vs-katago-kishin-ma...
It can't be just play like AI.
Any other Korean on the Korean Go program could have done the same.
In fact, many did when AlphaGo was the pinnacle of AI.
I wonder if this means the best Go play is closer to theoretically perfect play or if it just happened the current computer methods didn't manage to get much farther than humans. Go has vastly more valid games but also a simpler ruleset, so I'm not sure if there is really a good way to tell beyond "keep trying and find out"?
I recall a while back someone came up with a set of "anti-computer" strategies that allowed even an amateur to defeat a strong go program. These moves weren't anything like ordinary go moves (and perhaps the "loophole" has been closed now) but imo, their existence suggests that a study of programs may reveal other unexpected weakness.
Until a more average player confuses the AI with an unseen behaviour (pulling creeps between the towers etc) to get an advantage.
Anyway. Yes if you throw examples into training it will be able to handle the situation - but handling unseen things for me is a key goal.
This in no way detracts from how absurd and remarkable it is that Shin Jinseo can beat KataGo (it gets a LOT of training and architecture refinements https://katagotraining.org/#eloGraphButtons) with 2 stones of handicap.
Shin's genius was to play out a complex variation of the flying knife joseki that was, in essence, a one-way path to reach an equal board position that occupied about 1/4 of the board. Due to the 2-stone handicap, the position favoured black with the game ~25% complete. KataGo could not have played any other way, where a human may have tried to foil the plan by introducing further complications.
What was truly incredible was how Shin held the advantage from that point on.
Second, if it is a fixed sequence, how position independent is it? In chess, tactical sequences end up depending on the entire board state to work when they get sufficiently long. I guess what I'm asking is, could a player with some capacity for stategic thinking recognise this idea and take steps to make the flying knife impossible?
I really should spend some more time learning go, it's such a fascinating game.
However, when Shin executed the 50 move flying knife, the board was pretty much empty. So there is really no need for calculation, both Shin and the AI know it’s locally optimal. But getting to play a very long locally optimal sequence is good for the weaker player, so they have less “real” moves to lose EV on. Notably Shin probably can’t open with the flying knife in one corner past a certain point in the game, even if that corner were completely empty - the rest of the board positions would change the end values of the variants.
If the AI could know this, they might play a variant that ends 30 moves sooner but is 0.01 pts worse. Then they would have more time to mess Shin up through organic new moves (which the AI will be better at of course).
(disclaimer: only ranked 1 dan)
Shin took a 2-stone handicap from KataGo which means that Shin is the weaker of the two. But to give that more context, Shin is also the strongest human player to have ever lived in raw strength terms by a good margin, and is known as replicating AI move-for-move more closely than anyone else.
If they were to play even then there’s no chance any human could win (and pretty much all pros agree with that). Lee Sedol beating AlphaGo in game 4 of that series is largely considered the last time a human beat a modern AI in an even game, which is why it was so amazing.
RE the game, Katago was set to use the strongest available model and ran on a 3x 3090 GPU system, which is a lot for KataGo. 20 seconds might sound like a handicap, but that’s over 100,000 play out variations which is essentially infinite for modern KataGo models (anything over 10,000 is overkill).
Shin played well in all games, but his strategy was to avoid complexity. KataGo reads out complex fighting like an absolute monster, so Shin was trying to play very very solid and very very calm so as to not give KataGo an in.
The 2-stone handicap could be thought of as roughly 10-15 points of ‘buffer’. That’s massive in professional games, and that’s what Shin used to win. He played so overly solid that it sometimes cost a point or two, but it removed an opening for Katago to fight. He did this at the key opening and middle-game sections and never burnt through the full buffer of handicap points in the last two games. That’s why those games look kinda ‘boring’, it’s because Shin wanted them to be that way.
Also note that KataGo probably could have won if its ’variance’ was tuned higher (basically it taking risks). Standard KataGo won’t take risks, it just wins with brute force. For handicap games though you can tune its willingness to start fights higher to prevent people from just playing ultra solid (like Shin did).
Shin did an absolutely amazing job and he deserves all the recognition. Katago routinely beats professionals giving them 3-4 stones of handicap, so the win by Shin highlights to me how strong he is, but also just how well he understands how the AI ‘thinks’.
Also, the human played a strategy tailored to that huge initial advantage. He said that the AI did not handle this particularly well, and played high probability moves instead of trying to lure him into a mistake.
Also, even though this was the best Go engine, it was not running on a supercomputer, and had a relatively limited amount of time per move.
So, this was an important victory for a human, but not a sign that humans are now stronger than AIs at Go.
Yeah, katago's training is not really focused at all on handicap games, because it's by nature learning from even games against similar-strength opponents.
It doesn't have specific training from playing in a way to exploit a weaker player. In a handicap game you have to give your opponent opportunities to fuck up if you want to play optimally.
If a move loses 0.0005 points if the opponent plays optimally, katago won't play it even if there's ~zero chance a weaker player would play it right.
There have been go AIs that tried to train more directly on uneven opponents, one called "sai" comes to mind, but katago has huge advantages otherwise and won out over the others (for very good reason, it's a great project).
Just stating this off the top of my head so I could be misremembering, but I heard that the KataGo settings used were tweaked to favor complexity. This was most apparent in Game 1 which Shin Jinseo lost, where the AI had an unusual opening. However, the last game was quite plain leading me to wonder whether that setting was present in the last game (or at all).
You can kind of tweak towards play this metric or that, but it's not the same.
Another way to look at this: Go's handicap system gives us a genuinely interesting metric for the distance between a human and a machine at this specific game. Instead of just "computers beat humans" we get a quantified gap.
In any intellectual contest between human and machine, all the machine winning implies is that the endeavor is algorithmic.
The machine can be given practically unlimited memory and compute; we consider it cheating if the human would use memory aids. The machine could be implemented as many agents cooperating; we'd think it's not right if thousands of humans collaborated to face the machine, etc.
So statements like "not a sign that humans are now stronger than AIs at Go" are pretty meaningless, IMO
Totally wrongheaded, actually, since the computer gave the human a 2 stone advantage from the start.
It's never made sense to me that so many go players study AI go play in the hope of emulating it in a human game. We're not machines. We can't do thousands of Monte Carlo tree searches per second.
What a powerful story. Humans have a chance of remaining superior because emotions are the fuel for our intellect and wisdom.
Obviously its an impressive human intellectual feat but the hubris is instructive perhaps for our wider interactions with AI as its abilities accelerate around us.