[ExI] Google’s Go Victory Is Just a Glimpse of How Powerful AI Will Be

Anders Sandberg anders at aleph.se
Mon Feb 1 01:49:39 UTC 2016

Good point. There seems to be a fair bit of interest in DL hardware, and 
I think we will see ASICs as applications become standardized:

It is a one-time performance increment, but as you say maybe 3 orders of 
magnitude. Most obvious uses will likely be on the application side 
rather than the more expensive training side, but I think different 
ASICs would be able to boost both legs.

On 2016-01-31 21:52, Brian Atkins wrote:
> On 1/31/2016 1:04 PM, Anders Sandberg wrote:
>> More practically I think the Wired article gets things right: this is 
>> a big deal
>> commercially.
> Right. I find an interesting possible parallel with Bitcoin mining 
> hardware. Which obviously has a large economic incentive behind it. 
> Similar to Deep Learning tech, it started out as a CPU algorithm and 
> then eventually moved to GPUs. But then within a short period of 3 to 
> 4 years now has gone through a brief FPGA phase, and after that onto 
> sustained generations of custom ASIC designs that now have pretty much 
> caught up to state of the art Moore's Law (16nm designs being deployed).
> So I would expect something like possibly 3 orders of magnitude power 
> efficiency improvements could occur before 2020 as Deep Learning ASICs 
> start to deploy and get quickly improved? And more orders of magnitude 
> of total Deep Learning computing capability on top of that due to 
> simply deploying more hardware overall. Bitcoin mining just recently 
> crossed the exahash/s level... some of that due to ASIC improvements, 
> but a lot due to simply more chips being deployed. Deep Learning 
> takeoff could be an even quicker timeline than with Bitcoin since the 
> companies involved with Deep Learning stuff are much larger and better 
> funded.
> large disclaimer: I have no technical knowledge of whether there might 
> be some major difference between Bitcoin mining's hashing algorithm 
> and Deep Learning's typical algorithms that would make DL stuff 
> significantly harder to translate into efficient ASICs.
> _______________________________________________
> extropy-chat mailing list
> extropy-chat at lists.extropy.org
> http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat

Anders Sandberg
Future of Humanity Institute
Oxford Martin School
Oxford University

More information about the extropy-chat mailing list