To be fair, AlphaGo probably utilizes ensembles of networks in some form anyway. So, in a way, it's kind of been playing as a team this whole time :)
Google’s AlphaGo AI defeats team of five leading Go players
11–20 of 57 posts
Re: Google’s AlphaGo AI defeats team of five leading Go players
#12Earlier quoted context omitted.
With the caveat that it is a single, specific task. The team here of 5 is probably capable of more than playing Go :) That's not all bad, there are advantages to having something really good at one objective. But I wonder if there is a fundamental trade off.
IMO, you're comparing from the wrong things. AlphaGo is good at playing Go specifically, sure. But AlphaGo was built by machine learning infra that is capable of doing more than just playing Go. Google's ML capabilities as a whole are the 10x developer.
I've once read someone comparing the energy efficeincy of the human brain compared to all the CPUs used by one of these neural networks. The human brain turned out to be more efficient by many orders of magnitude.
Sadly I have no recollection of where I've read this.
Re: Google’s AlphaGo AI defeats team of five leading Go players
#13Earlier quoted context omitted.
IMO, you're comparing from the wrong things. AlphaGo is good at playing Go specifically, sure. But AlphaGo was built by machine learning infra that is capable of doing more than just playing Go. Google's ML capabilities as a whole are the 10x developer.
I've once read someone comparing the energy efficeincy of the human brain compared to all the CPUs used by one of these neural networks. The human brain turned out to be more efficient by many orders of magnitude. Sadly I have no recollection of where I've read this.
The brain is efficient but, let's be honest, not really productive, especially not 100% of the time.
Re: Google’s AlphaGo AI defeats team of five leading Go players
#14Earlier quoted context omitted.
IMO, you're comparing from the wrong things. AlphaGo is good at playing Go specifically, sure. But AlphaGo was built by machine learning infra that is capable of doing more than just playing Go. Google's ML capabilities as a whole are the 10x developer.
I've once read someone comparing the energy efficeincy of the human brain compared to all the CPUs used by one of these neural networks. The human brain turned out to be more efficient by many orders of magnitude. Sadly I have no recollection of where I've read this.
Even if it is, that's a questionably useful observation, because computers can consume and dissipate much more energy than the human brain.
Re: Google’s AlphaGo AI defeats team of five leading Go players
#15Earlier quoted context omitted.
IMO, you're comparing from the wrong things. AlphaGo is good at playing Go specifically, sure. But AlphaGo was built by machine learning infra that is capable of doing more than just playing Go. Google's ML capabilities as a whole are the 10x developer.
I've once read someone comparing the energy efficeincy of the human brain compared to all the CPUs used by one of these neural networks. The human brain turned out to be more efficient by many orders of magnitude. Sadly I have no recollection of where I've read this.
Luckily deep learning computers can directly use electricity produced from sources that don't compete with human nutrition or we'd be in for a very dark future.
Re: Google’s AlphaGo AI defeats team of five leading Go players
#16Is playing go as a team something that people do? It seems unintuitive. I'd expect a team of 5 experts to play worse than one expert.
At this point no one expect alphago would lose. So just have fun
I'd love to take this as far as possible.
What happens if you allow the team to roll back decisions as they see they're at a disadvantage? How far away are we from an effectively unbeatable machine?
Given what people can learn over time, can they learn to beat it?
Re: Google’s AlphaGo AI defeats team of five leading Go players
#17Algorithms are the new 10x developer. I'm actually half serious: one can see how something that is slightly better can have a massive multiplier, e.g. a team of inferior intelligences cannot 'scale' to beat a single superior intelligence in some regimes. This does not apply to all fields, of course. But in many regions of abstraction it may.
With the caveat that it is a single, specific task. The team here of 5 is probably capable of more than playing Go :) That's not all bad, there are advantages to having something really good at one objective. But I wonder if there is a fundamental trade off.
Great point. I'm inclined to agree, but we are all speculating. Right now, you are certainly right: we can craft algorithms that are super-human in certain respects, but they do not generalize well.
And of course, it took 100+ top PhDs in ML to create DeepMind. It certainly required many humans!
Re: Google’s AlphaGo AI defeats team of five leading Go players
#18Is playing go as a team something that people do? It seems unintuitive. I'd expect a team of 5 experts to play worse than one expert.
It is common. The Chinese commentator Gu Li mentioned pair game and team game. But I cannot find an English page about it. Here is a Chinese Wikipedia page. https://zh.m.wikipedia.org/zh-hans/团队围棋
FWIW, the five people team played with Ke Joe before the game and won (again, according to Gu Li.)
Re: Google’s AlphaGo AI defeats team of five leading Go players
#19Earlier quoted context omitted.
IMO, you're comparing from the wrong things. AlphaGo is good at playing Go specifically, sure. But AlphaGo was built by machine learning infra that is capable of doing more than just playing Go. Google's ML capabilities as a whole are the 10x developer.
I've once read someone comparing the energy efficeincy of the human brain compared to all the CPUs used by one of these neural networks. The human brain turned out to be more efficient by many orders of magnitude. Sadly I have no recollection of where I've read this.
This article claims that the current AlphaGo is running on just one TPU board:
https://www.wired.com/2017/05/googles-alphago-levels-board-g...
This article shows that there are 4 TPU2 accelerators per board and estimates TDP of 250 watts per accelerator:
https://www.nextplatform.com/2017/05/22/hood-googles-tpu2-ma...
Plus each of these boards has dual Xeon host processors. Maybe peaking at 1500-2000 watts all told per board, considering DRAM, storage, networking and power supply conversion losses? (I'm trying to be generous with the upper bounds.)
The human brain dissipates about 20 watts. But to date no game-playing-champion brains have been able to operate without the overhead of a host body attached to them. The basal metabolic rate of the human body is about 100 watts. That would make a human go player up to 20 times more energy efficient than a TPU board running AlphaGo (100 watts vs up to 2000).
Want to include all the energy that went into manufacturing the hardware, and the training phase? Don't forget to include the lifetime energy consumption of an adult human go player for parity.
It gets less favorable for the human with further analysis. AlphaGo can take on challengers tirelessly, 24/7. Human game players can play, what, 30 hours a week before losing their edge? Now the human is down to just 3.6x as energy-efficient as the machine; machines can fully power off while humans continue to dissipate significant power just sleeping.
The killer systemic disadvantage to the human side is that machines "eat" electricity while humans need food. The cheapest food energy sources, like potatoes, are far more expensive joule-for-joule than electricity. It also takes far more land area to grow a gigajoule of human-edible biomass per year than to produce and store a gigajoule of machine-usable electricity.
Science fiction stories sometimes portray far-distant futures where human and animal muscle power still perform menial tasks instead of machines because they consume "cheap" food instead of "expensive" electricity. The reality is that machines already have significantly lower operational costs than muscle power, even if you use fairly expensive electricity sources like battery-backed solar PV. They also have lower running costs for any thinking-like tasks they can actually perform. Add machines into the labor pool and the "wage" floor predicted by the Iron Law of Wages is too low to sustain human life. (Fortunately for humans, most countries set minimum wage floors by law rather than by unfettered market dynamics. But the long term trend of a shrinking percentage of humans able to do productive work more efficiently than machines will be... interesting.)
Re: Google’s AlphaGo AI defeats team of five leading Go players
#20To be fair, AlphaGo probably utilizes ensembles of networks in some form anyway. So, in a way, it's kind of been playing as a team this whole time :)
Well, what do you think your brain is? :)