News for hackers. Stuff that matters.
Re:Go master Cho wins best-of-three series against Japan-mad... by deegles on Thursday November 24, 2016 @10:05PM (#13034151)
What bothered me about AlphaGo was the sheer amount of power that was dedicated to training and running it vs. the relatively puny power budget a human has.

i.e. "Lee Sedol used about 20 Watts of power to operate. By contrast, AlphaGo runs on a whopping 1920 CPUs and another 280 GPUs for an estimated power consumption of approximately 1 MW (200 W per CPU and 200 W per GPU). That’s 50,000 times as much power as the amount of power that Lee Sedol’s brain uses and the two are not quite evenly matched but it is close enough to use for comparison."[0]

I mean, sure, it's apples to oranges, but it feels like a pyrrhic victory to have to use 1MW of computing power to beat a 20w brain.

I want there to be a series of AI vs. AI games where everyone has an equal power budget for their AI. Everyone has to bring their own equipment and they receive exactly X kwh from a calibrated power supply to train their model. Training data could possibly also be standardized.

Each team would received a standardized power supply for the actual competition (e.g 20w). There could be bonuses for using less energy for training or during gameplay, or for faster response times between moves.

I believe this would drive innovation in building special purpose hardware, more efficient training models, and so on.

[0] http://jacquesmattheij.com/another-way-of-looking-at-lee-sed...

Re:Go master Cho wins best-of-three series against Japan-mad... by mrob on Friday November 25, 2016 @02:54AM (#13034992)
Lee Sedol did not use 20W, because he is not a disembodied brain. A non-exercising human is usually approximated as using 100W.
Re:Go master Cho wins best-of-three series against Japan-mad... by monix on Friday November 25, 2016 @02:03AM (#13034827)
On the other hand AlphaGo proves that computers can scale, even if they are inefficient, whereas the only way to grow that 20 Watts brain would be to use multiple brains and that's not going to be any good in Go matches.
Re:Go master Cho wins best-of-three series against Japan-mad... by argonaut on Friday November 25, 2016 @06:40AM (#13035998)
On the contrary, I think there is a real question over whether AlphaGo would be able to beat a group of 3-5 tier-1 Go players (e.g. ~top 10 rankings players) who were allowed to write things down on paper, discuss, use whiteboards, use separate game boards to play variations, etc, even with same time controls.
Re:Go master Cho wins best-of-three series against Japan-mad... by arcticfox on Friday November 25, 2016 @01:55AM (#13034798)
why optimize for power? You can't create a megawatt human, that's our limitation that is being shown in these matches. power would just be a weird artificial constraint.
Re:Go master Cho wins best-of-three series against Japan-mad... by monix on Friday November 25, 2016 @02:07AM (#13034842)
Even though an argument can be made about computers being able to scale, energy is always the only constraint that matters.
Re:Go master Cho wins best-of-three series against Japan-mad... by Veratyr on Friday November 25, 2016 @03:51AM (#13035198)
> energy is always the only constraint that matters

Simply making a statement like this doesn't really say much. You really need to back it up with evidence, justification and/or reasoning, or at least elaborate on it a little.

Re:Go master Cho wins best-of-three series against Japan-mad... by rcthompson on Thursday November 24, 2016 @10:15PM (#13034203)
> Everyone has to bring their own equipment and they receive exactly X kwh from a calibrated power supply to train their model

What's the advantage of this over simply providing every team with identical hardware and equal computing time? It seems like providing a fixed power supply would favor whoever can afford to buy more efficient computing equipment.

Re:Go master Cho wins best-of-three series against Japan-mad... by deegles on Thursday November 24, 2016 @11:02PM (#13034348)
Developing special purpose machine learning hardware is part of the goal. I think of it like Formula 1, where the R&D eventually trickles down to consumer hardware.

Also, there's no reason why you couldn't have different competition categories where you're limited to using X dollars worth of time on a specific cloud provider, for example.

Re:Go master Cho wins best-of-three series against Japan-mad... by meric on Thursday November 24, 2016 @10:32PM (#13034253)
Measuring by power supply allows human participants if the limit is 20W.
Re:Go master Cho wins best-of-three series against Japan-mad... by TwoBit on Thursday November 24, 2016 @11:57PM (#13034488)
Identical hardware limits you.
Re:Go master Cho wins best-of-three series against Japan-mad... by shard on Friday November 25, 2016 @07:17PM (#13040908)
Another difference between the human brain and a computer is the number of round of hardware optimization. Human brains have millions of years and millions of search paths explored, while computer hardware, although advancing at a much faster pace, is still a relatively new technology.