Hacker News Reader

Comment by jacquesm | Hacker News Reader

With the amount of computing power they throw at this I doubt that's something even google could easily fund.

Is running AlphaGo really that expensive? I get that training deep learning systems is very computationally expensive, but my understanding is that running them is orders of magnitude cheaper.

edit: table showing gains as CPUs are added - https://i.imgur.com/xxdWUtV.png

It's a huge cluster of machines.

Got any numbers?

The Economist says "The version playing against Mr Lee uses 1,920 standard processor chips and 280 special ones developed originally to produce graphics for video games"

So, that's actually not very expensive to rent for the duration of five games.

I don't know how long a game takes but I don't think google could afford letting anybody play, as stated in the original post.

Charge $1000 per game and pay out $10,000 if the human wins. Could be a nice earner. ;)

I'll pay for two simultaneous sessions of AlphaGo please :). I'll even let it go first in one of them!

Clever, but they could easily detect this on their servers.

It would be nice if they preserve a snap shot of Alphago as it was a the start of the Sedol games as a historic thing. Then if they open source it people could go play it too. They have mostly open sourced their code but not I think the learning data.

They would only need to open the learned coefficients, not the data learned from.

I like the way you think. :D

If you price by the GCE calculator[0] it's $1920 to rent 1920 CPU cores for 20 hours. This doesn't include GPU costs as they don't seem to have GPUs available on cloud, but I could see that easily doubling the costs.

[0] https://cloud.google.com/products/calculator/#id=f63f1fe0-56...

Daresay the Economist is unclear whether they are talking chips or cores so it could be several times more cores.

But that would prove that they are not overfitting their opponent. Which would be nice.

Not really. Even if it would only take a single dedicated GPU to win in real time against a human opponent you could not offer a service like that for free. If google massively overfits their opponent that equation still holds true so it's not really proof. It would only be proof if they switched off 10% or so of the array that powers their offering right now and it would lose from Lee Sedol consistently. For all you know they have a huge margin of error.

I suspect that training the AI is the most CPU intensive part.

If they would release the (trained) AI to everybody, that would prove that the training phase is general enough to beat any player, not just one.

Google doesn't have to pay for the CPU time. Ke Jie can find sponsors if he pretends he can beat AlphaGo.

(Edit: clarify)

I don't doubt there is an a-symmetry between the number of machines / GPUs they throw at the problem during a match and during the run-up to a match but even so they will have to have some margin of error if they expect to win in the first place and besides that whatever that pile of hardware is it, the infrastructure required to run it and the people involved are not free.

Alas, the software is tied to Google's platform. Just releasing as open source won't do. (That's why they didn't open source Google Reader. The source would have been both useless for running a Reader clone, and given away lots of trade secrets about Google infrastructure.)

Isn't one of the features of AlphaGo that it trains one part of the AI while playing the game?

Reply on news.ycombinator.com