AlphaZero Engine Availble to Public?

Sort:
Master_Po

Anyone heard any news about AZ being available to the public?  Am sure it'll be expensive.  

notmtwain
Master_Po wrote:

Anyone heard any news about AZ being available to the public?  Am sure it'll be expensive.  

Sure, just buy a controlling interest in Google.

liml

I don't know what you intend to do with an AZ engine because it runs on a very specialise hardware called TPU (Tensor Processing Unit). Our desktops and laptops have a CPU. The TPU is quite a behemoth. It's capable of 45 teraflops (TF). The latest and greatest Intel CPU is capable of 1 TF. See this discussion for more details: https://chess.stackexchange.com/questions/19366/hardware-used-in-alphazero-vs-stockfish-match

mikecantreed1

Lol Stockfish 8 too easy for you?

EscherehcsE
Master_Po wrote:

Anyone heard any news about AZ being available to the public?  Am sure it'll be expensive.  

Not AZ itself, because of the specialized hardware as mentioned previously. (And Google may not want to share the knowledge, anyway.) However, there is already an open source volunteer effort underway to reproduce the GO part of AZ. It's called Leela Zero:

http://talkchess.com/forum/viewtopic.php?t=65481&start=40

 

Also, there's a more recent attempt just started by another person to copy the Leela Zero methods and apply them to chess. It's called Leela Chess Zero (or lczero, for short):

http://talkchess.com/forum/viewtopic.php?t=66280

Master_Po

Thanks, liml and EscherehcsE.   

bladesclawsfangs

Probably not. Deepmind/Google is in the machine learning and AI business, not producing a chess product for consumers. 

Godeka

AlphaZero is no product that will be sold. It‘s on the chess community to develop an AlphaZero-like engine. Maybe the code base of LeelaZero can be used? Wouldn’t it be very nice to have a multi-game engine for games like chess, Go, Shōgi, Xiangqi?

 

@liml:

Even if there are NN engines can be fairly strong when using the CPU only, usually graphic cards are used fort he calculations because the are much faster for massive parallel calculations. It’s not necessary to have tensor cores. And tensor cores will become cheaper in the next years.

tenkeren

Tensor cores are available for rent from Google via an online interface, so there is no need to buy the hardware. See this link https://cloudplatform.googleblog.com/2018/02/Cloud-TPU-machine-learning-accelerators-now-available-in-beta.html


By buildling a web interface on top of this, AlphaZero could be made available as a paid service. The service could be built by a third party (professional) chess enthusiasts in collaboration with Google, and the latter would not have to disclose the source code of AlphaZero or deliver anything else than an API endpoint for the web service to communicate with through the UCI protocol.

The question is if the market is big enough to break even on cost, and of course, if Google are willing to collaborate on such a project.

I can understand people lightly sarcastic asking if Stockfish is not a strong enough opponent, but this is not all about ELOs, it is the playing style of AlphaZero that would be of interest to experience, described as human-like and even super-human.