<div dir="ltr">Thanks for clarifying the TPU counts, I should have read the paper more carefully. That changes what I said earlier a great deal in terms of computing horsepower used for training and game simulation. They actually used an absolutely tremendous amount of computing power that is not easily available to the general populace. This doesn't take away from the accomplishment, but it explains the training time being accomplished in less than half a day.<br></div><div class="gmail_extra"><br><div class="gmail_quote">On Fri, Dec 8, 2017 at 3:35 AM, Alejandro Dubrovsky <span dir="ltr"><<a href="mailto:alito@organicrobot.com" target="_blank">alito@organicrobot.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">The 4TPUs were what was used during playing time. During training they say it used 64 second-generation TPUs for "training the neural network" ie the gradient, backprop etc of the neural network itself. It also used another 5,000 first-generation TPUs to generate the games. Both of those are from page 4 of their paper. I've got no idea of the computing power of those TPUs.<br>
<br>
BTW, while this is, to me, the biggest newest in the history of chess, it follows naturally from their AlphaGo Zero paper published on Nature a whole month and a half ago: <a href="https://www.nature.com/articles/nature24270" rel="noreferrer" target="_blank">https://www.nature.com/article<wbr>s/nature24270</a><br>
<br>
I don't think there's anything sinister in the whole affair, except for the usual DeepMind showboating (and the usual question of why they denied StockFish their opening book and only gave it 1GB of hashtable memory). I expect people to indepedently replicate this, albeit at a slower pace due to limited resources, within the next couple of years.<span class=""><br>
<br>
<br>
On 08/12/17 07:10, Dylan Distasio wrote:<br>
</span><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><span class="">
The paper describes in detail how they did it and how long it took. Again, I see no reason to question any of it. It's an evolutionary advance from their work on Go, and is using established deep learning and reinforcement techniques on powerful hardware (although not insanely powerful, 4 TPUs is not unobtainium <a href="https://www.blog.google/topics/google-cloud/google-cloud-offer-tpus-machine-learning/" rel="noreferrer" target="_blank">https://www.blog.google/topics<wbr>/google-cloud/google-cloud-<wbr>offer-tpus-machine-learning/</a>).<br>
<br>
<br></span><span class="">
On Thu, Dec 7, 2017 at 3:01 PM, Tomaz Kristan <<a href="mailto:protokol2020@gmail.com" target="_blank">protokol2020@gmail.com</a> <mailto:<a href="mailto:protokol2020@gmail.com" target="_blank">protokol2020@gmail.com</a><wbr>>> wrote:<br>
<br>
Like spike, I also think something hasn't been told or at least<br>
something hasn't been emphasized as it should be.<br>
<br>
But if it's not today, it will be in a year or two. Doesn't really<br>
matter. (What does matter, is that the Singularity is closer than ever.)<br>
<br>
On Thu, Dec 7, 2017 at 8:34 PM, Dylan Distasio <<a href="mailto:interzone@gmail.com" target="_blank">interzone@gmail.com</a><br></span><span class="">
<mailto:<a href="mailto:interzone@gmail.com" target="_blank">interzone@gmail.com</a>>> wrote:<br>
<br>
I think you would be surprised at how little code is actually<br>
there if we could see the source code. Tensorflow does a lot of<br>
heavy lifting in terms of abstracting neural nets, and the<br>
secret sauce in the recipe is the trained weights of the net. There is very little conventional code beyond setting up the<br>
structure of the neural net and the reinforcement objectives. Training on 77 million simulated games is the key.<br>
<br>
On Thu, Dec 7, 2017 at 2:15 PM, spike <<a href="mailto:spike66@att.net" target="_blank">spike66@att.net</a><br></span>
<mailto:<a href="mailto:spike66@att.net" target="_blank">spike66@att.net</a>>> wrote:<br>
<br>
__ __<br>
<br>
__ __<br>
<br>
*From:* spike [mailto:<a href="mailto:spike66@att.net" target="_blank">spike66@att.net</a> <mailto:<a href="mailto:spike66@att.net" target="_blank">spike66@att.net</a>>]<br>
*Sent:* Thursday, December 07, 2017 10:53 AM<br>
*To:* 'ExI chat list' <<a href="mailto:extropy-chat@lists.extropy.org" target="_blank">extropy-chat@lists.extropy.or<wbr>g</a><br>
<mailto:<a href="mailto:extropy-chat@lists.extropy.org" target="_blank">extropy-chat@lists.ext<wbr>ropy.org</a>>><br>
*Subject:* RE: [ExI] alpha zero____<br>
<br>
__ __<br>
<br>
__ __<br>
<br>
__ __<br>
<br>
*From:* extropy-chat<br>
[mailto:<a href="mailto:extropy-chat-bounces@lists.extropy.org" target="_blank">extropy-chat-bounces@l<wbr>ists.extropy.org</a><br>
<mailto:<a href="mailto:extropy-chat-bounces@lists.extropy.org" target="_blank">extropy-chat-bounces@l<wbr>ists.extropy.org</a>>] *On Behalf<br>
Of *spike<br>
*Subject:* Re: [ExI] alpha zero____<br>
<br>
__ __<br>
<br>
__ __<br>
<br>
__ __<br>
<br>
*From:* extropy-chat<br>
[mailto:<a href="mailto:extropy-chat-bounces@lists.extropy.org" target="_blank">extropy-chat-bounces@l<wbr>ists.extropy.org</a><br>
<mailto:<a href="mailto:extropy-chat-bounces@lists.extropy.org" target="_blank">extropy-chat-bounces@l<wbr>ists.extropy.org</a>>] *On Behalf<br>
Of *John Clark____<br>
<br>
__ __<br>
<br>
>>>…I still haven’t convinced myself it is true. spike____<br>
<br>
__ __<br>
<br>
__ __<br>
<br>
__ __<br>
<br>
>>…If this is a hoax it's a very elaborate one the likes of which we haven't seen since the cold fusion fiasco. …John K Clark ____<br>
<br>
__ __<br>
<br>
__ __<br>
<br>
__ __<span class=""><br>
<br>
>…Ja I could have clarified my doubt a bit. I don’t<br>
suspect an intentional hoax, rather something they neglected<br></span>
to tell us…spike____<br>
<br>
__ __<br>
<br>
__ __<br>
<br>
__ __<span class=""><br>
<br>
There is something else. If someone writes a chess-playing<br>
program today that defeats the best effort of the best<br>
humans for 500 years, that has been done. According to the<br>
article, what Alpha Zero has done is defeat the collective<br>
programming efforts of armies of professional programmers<br>
working with huge monetary rewards for the past 75 years,<br>
which is far more impressive than overpowering humans in<br></span>
chess.____<br>
<br>
__ __<br>
<br>
Something is still not being told here methinks.____<br>
<br>
__ __<br>
<br>
spike____<span class=""><br>
<br>
<br>
______________________________<wbr>_________________<br>
extropy-chat mailing list<br>
<a href="mailto:extropy-chat@lists.extropy.org" target="_blank">extropy-chat@lists.extropy.org</a><br></span>
<mailto:<a href="mailto:extropy-chat@lists.extropy.org" target="_blank">extropy-chat@lists.ext<wbr>ropy.org</a>><br>
<a href="http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat" rel="noreferrer" target="_blank">http://lists.extropy.org/mailm<wbr>an/listinfo.cgi/extropy-chat</a><span class=""><br>
<<a href="http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat" rel="noreferrer" target="_blank">http://lists.extropy.org/mail<wbr>man/listinfo.cgi/extropy-chat</a>><br>
<br>
<br>
<br>
______________________________<wbr>_________________<br>
extropy-chat mailing list<br>
<a href="mailto:extropy-chat@lists.extropy.org" target="_blank">extropy-chat@lists.extropy.org</a><br></span>
<mailto:<a href="mailto:extropy-chat@lists.extropy.org" target="_blank">extropy-chat@lists.ext<wbr>ropy.org</a>><br>
<a href="http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat" rel="noreferrer" target="_blank">http://lists.extropy.org/mailm<wbr>an/listinfo.cgi/extropy-chat</a><span class=""><br>
<<a href="http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat" rel="noreferrer" target="_blank">http://lists.extropy.org/mail<wbr>man/listinfo.cgi/extropy-chat</a>><br>
<br>
<br>
<br>
<br>
-- <a href="https://protokol2020.wordpress.com/" rel="noreferrer" target="_blank">https://protokol2020.wordpres<wbr>s.com/</a><br>
<<a href="https://protokol2020.wordpress.com/" rel="noreferrer" target="_blank">https://protokol2020.wordpres<wbr>s.com/</a>><br>
<br>
______________________________<wbr>_________________<br>
extropy-chat mailing list<br></span>
<a href="mailto:extropy-chat@lists.extropy.org" target="_blank">extropy-chat@lists.extropy.org</a> <mailto:<a href="mailto:extropy-chat@lists.extropy.org" target="_blank">extropy-chat@lists.ext<wbr>ropy.org</a>><br>
<a href="http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat" rel="noreferrer" target="_blank">http://lists.extropy.org/mailm<wbr>an/listinfo.cgi/extropy-chat</a><span class=""><br>
<<a href="http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat" rel="noreferrer" target="_blank">http://lists.extropy.org/mail<wbr>man/listinfo.cgi/extropy-chat</a>><br>
<br>
<br>
<br>
<br>
______________________________<wbr>_________________<br>
extropy-chat mailing list<br>
<a href="mailto:extropy-chat@lists.extropy.org" target="_blank">extropy-chat@lists.extropy.org</a><br>
<a href="http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat" rel="noreferrer" target="_blank">http://lists.extropy.org/mailm<wbr>an/listinfo.cgi/extropy-chat</a><br>
<br>
</span></blockquote><div class="HOEnZb"><div class="h5">
<br>
______________________________<wbr>_________________<br>
extropy-chat mailing list<br>
<a href="mailto:extropy-chat@lists.extropy.org" target="_blank">extropy-chat@lists.extropy.org</a><br>
<a href="http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat" rel="noreferrer" target="_blank">http://lists.extropy.org/mailm<wbr>an/listinfo.cgi/extropy-chat</a><br>
</div></div></blockquote></div><br></div>