Tweet from Demis Hassabis (CEO of DeepMind):

How many stones does it plan on giving Iyama?

_________________
Still officially AGA 5d but I play so irregularly these days that I am probably only 3d or 4d over the board (but hopefully still 5d in terms of knowledge, theory and the ability to contribute).

How about this wild specuation: In a few days, Zen will announce Zen vs Iyama. Zen wins.

The AlphaGo announcement is that it will participate in the UEC Cup -- a computer vs computer competition in Japan usually held in March. This becomes the new strongest tournament in the world.


ETA:

I can't help feeling that because AlphaGo's public debut was such a big event with Lee Sedol that they feel the need for subsequent events to be of a similar stature. So we're unlikely to see it play on KGS for example (particularly with only a medium-powered server) because if it loses Google/DeepMind would lose a lot of the highly valuable 'awesome at AI' brand they have built for themselves. So how about some kind of 10-game match against the world's top pros (either in series or consultation) where if it wins 2 or 3 games in a row the handicap goes up.

Quite reasonable; like what Go Seigen did with his contemporaries in 10-game series.

OTOH, if '3' is the threshold, then Takao would've pushed Iyama down to 1-stone,
and then Iyama would've climbed back to even in their latest Meijin sequence.

Yes, losing 3 games in a row doesn't really mean you are 1 stone weaker, but I erred on the smaller side so that within a 10 game match AlphaGo could have the chance to demonstrate if it is 3 stones stronger than top pros (actually if it is you would only expect something like 70% win rate on 2 stones so not winning 3 in a row on 2 stones is quite likely). And I think you'd need to at least start on even to not hurt the humans' pride too much .

Yes. Based on everything said and published by Mr. Hassabis,
Mr. Huang, and all of the DM team, they've always shown great respect
to Go and to the pros they've worked with. This is particularly significant
as we all know Go just happens to be a stepping stone
(after the classic ATARI games, and before StarCraft II) for their research.
No doubt they'll continue to extend the same courtesy to Blizzard
and to the SC2 pros.

Looking forward to 2-4+ stones to the top pros.
( On a device the size of a wrist watch. )

They could do exactly the same as Lee Sedol vs. American gopro handicap matches a few years ago to avoid any controversies regarding respect to gopro. But at least in the next two days we'll know if zen will also involve or not. Zen could be a guest program to take on an even event while AlphaGo is the main show, handicap event. Or maybe Zen vs. AlphaGo.

Aja Huang says that AlphaGo is significantly stronger than 6 months ago (unsurprising, but there it is for confirmation).

I'd love to see them go with some kind of hardware limitation. Single computer, 8 cores, 32 GB RAM. Something you still buy as a consumer, if you really wanted to.

I don't share this desire to limit AlphaGo (either by hardware of making it play 10 second blitz as was suggested elsewhere): it's like you've built a rocket and then you put a 10mph speed limiter on it. I don't care about throttling it down to human levels, if it can blast Ke Jie of the board and give him handicaps I want to see it! For sure how strong it is on low-spec hardware is interesting and relevant for future mass-use (and I expect it'll be similar to chess), but for a big exhibition match (if that's what's going to happen) crank up the TPUs and full-steam ahead!

Edit: and Aja, on the advice of his colleagues, just edited his post from "significantly stronger" to "massively stronger".

Me, too!



4 stones! 4 stones! 4 stones!

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

Both would be great. Four stones against Ke Jie on massive hardware, and then even games on consumer-grade hardware.

For big event I want to see how far AlphaGo can go without hardware limitation.
But next year (I hope) the single machine version of AlphaGo will play on KGS.
Even though DeepMind's definition of single machine is very demanding.... 48 CPUs and 8 GPUs (but some hardcore gamers are probably have PC that can compete with this)

Aja Huang's change from significantly to massively very surprised me because he never said something like this, even the big jump from v.13 to v.18. But look like DeepMind will have to visit Korea many times in the future for sure since their other game AI project is also very big hit in that country.

I'm also really surpised with the change of wording to "massively" because of the reputational importance of AlphaGo to DeepMind that Uberdude mentioned earlier.

The weight of expectation on AlphaGo now will be enormous.

To expand on Uberdude’s comment, Aja Huang’s comment on his — edited — post on FB:



Automatic translation on FB makes this of it:

So… something like … geologist’s hammer vs. steam hammer?

_________________
“The only difference between me and a madman is that I’m not mad.” — Salvador Dali ★ Play a slooooow correspondence game with me on OGS?

I have to wonder in what basic this claim is made. There isn't anyone massively stronger to compare it to. Did it play and earlier version of itself to come to this conclusion?

_________________
Still officially AGA 5d but I play so irregularly these days that I am probably only 3d or 4d over the board (but hopefully still 5d in terms of knowledge, theory and the ability to contribute).

This part seems relatively trivial: handicap stones. Either against previous versions, or against human pro(s).

I was actually assuming that had it played again another pro we would have heard about it, but perhaps that is not so.

_________________
Still officially AGA 5d but I play so irregularly these days that I am probably only 3d or 4d over the board (but hopefully still 5d in terms of knowledge, theory and the ability to contribute).

(Good) Scientists tend to avoid the word "significantly" in publications as it might be read as "statistical significant", which requires significantly more work to show (pun intended).

But I think Aja's comment means something else, as it was not changed as if fixing a typo.