[ExI] alpha zero

[Alejandro Dubrovsky]

On 11/12/17 02:42, John Clark wrote:
> 
> On Fri, Dec 8, 2017 at 3:35 AM, Alejandro Dubrovsky 
> <alito at organicrobot.com <mailto:alito at organicrobot.com>> wrote:
> 
>     ​> ​
>     The 4TPUs were what was used during playing time.
> 
> ​And ​
> Stockfish
> ​used​
>   64 CPU cores
> ​.​
> 

Yes. It's hard tell how the two compare. My guess is that the 4TPUs are 
doing way more FLOPs than the 64 cores. Just one of the latest TITAN Vs 
from Nvidia claims to do 110TF. There's no Intel beast going much past 
1TF, and each one of those has 24 CPUs, so max around 3TF in the 64 CPU 
cores. I think the assumption that a TPU designed to run neural-network 
code can at least keep up with a GPU at neural-network specific code is 
reasonable. All of that is a bit misleading since it's much easier to 
make close to full use of a CPU than it is of a GPU.

Not that any of that matters much really. Firstly, StockFish can't make 
use of GPUs/TPUs nor can any of the other engines. They have no use for 
FLOPs at all since it's all integer work.

More importantly, I don't think it really matters whether it can beat 
StockFish/Houdini/Komodo/whatever or whether it gets close. The 
impressiveness would diminish only very slightly in my eyes if it just 
got close (although I can see that it'd make a difference to others, and 
this is where the showboating comes in and where denying StockFish an 
opening book that it normally uses and this leading to those positions 
where StockFish is all tangled up due to a bad opening, just so that 
they can claim that it crushed StockFish and show off those strangling 
positions does make it a bit mischievious in my view). If all they'd 
done is learn to play chess up to around the level of StockFish 
completely by self-play in a short amount of time, using a generic 
algorithm which they also used to learn Go and Shogi, it'd still be huge 
news, and I completely believe that they've done that.