World #1 Go Player Ke Jie accepts Google Alpha Go Match..

[Milos]

No, if you follow "Nature" paper, 60-70% (especially tactically, MCTS) of gain comes from CPU.

Absolutely wrong. Why is it so hard to see the obvious from the paper???
It is enough just to look at Extended Data Table 10 (Scalability study).
2 threads vs 1 thread 70% winning chance.
2 GPUs vs 1 GPU 95% winning chance.

[duncan]

It is obvious they don't want to embarrass Ke Jie, but instead make it look as competitive as possible and in the same time still win 3:0. I wouldn't be surprised if this was even written in some contract.

how could they do that without weakening the software.?

[Milos]

It is obvious they don't want to embarrass Ke Jie, but instead make it look as competitive as possible and in the same time still win 3:0. I wouldn't be surprised if this was even written in some contract.

how could they do that without weakening the software.?

Exactly as Kai explained by the optimization function. By maximizing winning percentage but completely ignoring points difference (so instead of 2 parameters they only have 1 in their optimization function). So effectively their system is suboptimal.

[Laskos]

No, if you follow "Nature" paper, 60-70% (especially tactically, MCTS) of gain comes from CPU.

Absolutely wrong. Why is it so hard to see the obvious from the paper???
It is enough just to look at Extended Data Table 10 (Scalability study).
2 threads vs 1 thread 70% winning chance.
2 GPUs vs 1 GPU 95% winning chance.

Not that wrong, when I see the table, improvements from CPU and GPU come pretty equal, about 700 ELO points each. GPU improvement dampens faster. So, I stand corrected in the sense that 50% comes from CPU, 50% from GPU, in their reasonable range (1->40 threads, 1->8 GPUs). First doubling is indeed large for GPU, that's why I have 2 of them for a 4 core machine.

[Milos]

No, if you follow "Nature" paper, 60-70% (especially tactically, MCTS) of gain comes from CPU.

Absolutely wrong. Why is it so hard to see the obvious from the paper???
It is enough just to look at Extended Data Table 10 (Scalability study).
2 threads vs 1 thread 70% winning chance.
2 GPUs vs 1 GPU 95% winning chance.

Not that wrong, when I see the table, improvements from CPU and GPU come pretty equal, about 700 ELO points each. GPU improvement dampens faster. So, I stand corrected in the sense that 50% comes from CPU, 50% from GPU, in their reasonable range (1->40 threads, 1->8 GPUs). First doubling is indeed large for GPU, that's why I have 2 of them for a 4 core machine.

Not really, at least not in case valid for smartphones. 50% comes only when system is well balanced and has a lot of GPU power (bear in mind GPUs in Lee Sedol match were those custom extremely powerful TPUs).
40 threads 1 GPUs is only 18% vs. 4 threads, 8 GPUs. So 10 times more cores is much, much weaker (in terms of Elo particularly since 1% advantage is much more Elo than in chess) than 8 times more GPU.
When GPU power is low, the difference in favour of GPU strength is really pronounced. And this is particularly strong point in case of smartphones which have pretty low GPU power compared to CPU power.

[Laskos]

Game 2 pretty much decided by move 80 according to Crazy Stone, on the lower left corner. 11.5 points advantage, 80% probability of Win for Black (AlphaGo). "Mistakes" by Ke Jie moves 44 and 76.

[whereagles]

and another sweep for the silicon overlord

[Jeroen]

AlphaGo also won the 2nd game, as expected.

Although I have very limited Go knowledge, it looks to me that the opening is the Achilles heel of the top players and the phase where they get outplayed by AlphaGo.

[Jeroen]

Edit: in the post mortem team AlphaGo said that Ke Jie played perfectly in the first 50 moves (perfectly in a sense that AlphaGo predicted all Ke Jie's moves) and that even after 100 moves it was not yet clear that AlphaGo would win. So perhaps it wasn't such a onesided game after all.

[Laskos]

Edit: in the post mortem team AlphaGo said that Ke Jie played perfectly in the first 50 moves (perfectly in a sense that AlphaGo predicted all Ke Jie's moves) and that even after 100 moves it was not yet clear that AlphaGo would win. So perhaps it wasn't such a onesided game after all.

The DeepMind team suffers from fake super-politeness, as though they have to apologize for their wins. First 50 moves were indeed played well by Ke Jie, and the ensuing fights gave the game a volatile score (error margins increased instead of decreasing). But by move 80, as I posted earlier, the game was almost decided. This is according to my Crazy Stone Deep Learning 7d. Here is the Crazy Stone 2 hour analysis:

AlphaGo advantage in points:


Likelihood of AlphaGo Win (t-value):