I am curious to know what other people think. Is AlphaGo going to disappear as quickly as it arrived? There seems to be indications that the development team are interested in using it as a stepping stone to much more complex AI problems. At the same time I have heard of some interest in pursuing the program in the go world.

I just don't know what to expect. Is it going to continue to beat high level pros? Is the honorary 9 dan pro certificate pointless as it will not be playing this time next year? Are other programs going to supersede it because it will not continue to be developed and they will? Or is AlphaGo going to stick around and start giving top pros a handicap?

Or is it going to become Kami no Itte?

These, and many other questions, come to my mind. But I have no answers. What do you all think?

_________________
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).

I heard a rumor that an offer/invitation had already been extended to DM for Mr. Iyama to play AG.

I also don't know what's going to happen; as you mentioned, we all know Go is "merely" a stepping stone for DM
( along with earlier stepping stones, the Atari games ).

As Mr. Demis Hassabis said in the post-Game 5 Q&A, their already published the paper in Nature;
anyone is welcome to use the info to build their own version of AG,
and in fact, some people have already started -- Demis said this much.

My hunch is that DM will not pursue Go much further -- it's very likely they'll analyze and hunt down
whatever bugs or weakness exposed in Game 4 ( this is important not for Go, which they don't care so much
now that they've already achieved the big milestone of beating a top human pro -- unthinkable
even just a year ago -- but for all their future projects ). They may continue to start a project
to train AG ( or BetaGo, etc. ) from "scratch", instead of starting from high-dan or pro games.
If BG opens at tengen, it'll be so exciting.

They may host another, smaller event -- say if Mr. Iyama v. AG/BG happens --
but they will not spend all the energy and efforts like this AG-Lee Sedol event.

They'll move on. ( Other people -- newspapers, other media companies, noodle companies, mattress companies --
can put up the money and resources to make their own version of human-computer Go events for eyeballs. But not Google. )
That's my guess.

We don't need to rely on AlphaGo on challenging pro, other programs are already on the way, Deep Zen Go may take on Iyama Yuta while AMD+Baidu AI is already challenging Ke Jie (which he turn them down cause Baidu AI hasn't beat any othet pro yet)

Good luck training your neural networks without Google's infrastructure.



I agree. This is part of the reason that it was somewhat sad to me that AlphaGo beat Lee Sedol in the match. To people like Lee Sedol - even to people like me - Go is very important.

Demis Hassabis and his team are great at what they do. But deep learning is their #1 passion - not Go. That's why it would have been happier for me if Lee Sedol had won the contest soundly - at least for this challenge. They may have made a stronger program and beat him later, but at least it would say something toward the devotion that people put toward this game.

For Lee Sedol - for me - and for many other Go players... We'll continue to study Go. Naturally, Deep Mind will continue to pursue deep learning.

Can't help but feel a little inferior in this regard, though. They'll continue to solve the world's problems, while we remain studying on a problem that they've already shown their superiority in. Nonetheless, I feel a little bit better in knowing that, even if we're pursuing a problem they've already shown superiority in, it's still an interesting problem worth studying and worth improving ourselves in.

_________________
be immersed

Nobody said it would be easy, or cheap, either.
If it were, it wouldn't be worth $1,000,000, or it wouldn't have surprised so many who'd thought this feat was at least a decade away, would it.
But DM made their breakthrough approach public.

I agree. Just saying that it's not likely that an AI as strong as theirs will be coming along anytime soon even if the approach is known, because the computing resources required for the training is not feasible for many right now.

This is something that could have only happened from a company like Google that has that power.

_________________
be immersed

I can't find the source, I think it might have been something Andrew Jackson said, but supposedly the non-distributed version of AlphaGo won 30% of its games against the full distributed version. If that is true, then maybe smaller-scale operations have a shot!

_________________
- Brady
Want to see videos of low-dan mistakes and what to learn from them? Brady's Blunders

Well, I believe the CEO said the distributed version won 75% of the time.

Anyway, this is not the point, because the neural networks have already been trained. What is resource intensive is not playing the games, but the training.

Nobody could have the computation power to train that quickly but Google - a smaller company could not do this. After it is trained, there is less required computation power.

_________________
be immersed

Well Google went to Korea hoping to find a bug in their AI. So they walk away doubly happy. I hope they make a version 19 (Think they said current version was 18 and from the screen shot the front end looks like it was running on Unity on Ubuntu).

They said they have further challenges for AlphaGo (would be nice if they played a top pro from each country!) and some other projects over the summer. So it won't be dissappearing over night.

When they are done and dusted I hope they provide some more information to the community as I'm sure there are still some very important unanswered questions.

I wonder if we could try and do a seti@home thing and ask interested people (computers) to participate. These things seem to have lost some popularity recently, I'm guessing partly over security fears and partly how easy and affordable it now is to run 1000s of instances at an Amazon or Google, but there may still be a place for it.

I don't think the team is quite done with AlphaGo yet, but I doubt it will be their primary development tested for much longer. I think there will be a few more interesting data points for the go community, though. In particular, the team said they were going to make a version that learns to play without being primed by human games. This could potentially lead to a bot that is just as strong (or stronger!), but with a completely alien play style. Or it could converge upon a recognizably human style of play on its own (though I doubt it.) I'm curious what will come of that.

The problems with training from scratch is :
- it will probably take far more longer to get to the same playing strength
- it may stumble on a "local optimum" during learning

That is a STANDARD problem in training a neural net to do anything. Read the literature about neural nets to learn how "local maxima" are overcome. The process of training a neural net isn't one of linear improvement. More often train (improve) and then damage, repeat, repeat, repeat.

I would be more worried that the neural network might have found the same local maximum that humans are stuck at, than fearing that it might get stuck in a different maximum. At least the second one would be more interesting .