6 ms·

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

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.

dullgiulio 9y ago
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.
- jrimbault 9y ago
  The brain is efficient but, let's be honest, not really productive, especially not 100% of the time.
- obstinate 9y ago
  Even if it is, that's a questionably useful observation, because computers can consume and dissipate much more energy than the human brain.
- narrator 9y ago
  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.
  - daveguy 9y ago
    You say that now, but then it installs a Dyson sphere. Very dark indeed.
- philipkglass 9y ago
  This article claims that the current AlphaGo is running on just one TPU board: https://www.wired.com/2017/05/googles-alphago-levels-board-games-power-grids/ 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-machine-learning-clusters/ 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.)
- blahbwah777 9y ago
  Sure, but I wasn't making any rigorous statements around human brain power vs Google ML I was suggesting more simply that the original comment talking about 10x developer wasn't referring to just AlphaGo. Apparently it wasn't clear that I'm limiting my comments strictly to what's observable in the chat forum and not making sweeping statements about humans. Though I suspect self-selection bias might also explain the downvotes. Of course everyone on HN is better than a machine, right? And the brain itself may be. But is the average developer?
  - nightski 9y ago
    I didn't down-vote you, but I do disagree. This isn't some desperate attempt to hold on to some sort of artificial superiority to machines. I'd be thrilled if machines could do my job, because honestly half of it is mundane (the other half is communicating and interacting with humans). I'm just being honest about the limitations of current models. I'm also very interested about the impact of generalization. If we make a single model capable of multiple tasks, will that sacrifice the ability to perform on each individual task? Or will it scale?
- kingvash 9y ago
  Not a huge fan of the article but here it is: https://jacquesmattheij.com/another-way-of-looking-at-lee-sedol-vs-alphago https://jacquesmattheij.com/another-way-of-looking-at-lee-se...

dullgiulio 9y ago

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.

jrimbault 9y ago
The brain is efficient but, let's be honest, not really productive, especially not 100% of the time.
obstinate 9y ago
Even if it is, that's a questionably useful observation, because computers can consume and dissipate much more energy than the human brain.
narrator 9y ago
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.
- daveguy 9y ago
  You say that now, but then it installs a Dyson sphere. Very dark indeed.
philipkglass 9y ago
This article claims that the current AlphaGo is running on just one TPU board: https://www.wired.com/2017/05/googles-alphago-levels-board-games-power-grids/ 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-machine-learning-clusters/ 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.)
blahbwah777 9y ago
Sure, but I wasn't making any rigorous statements around human brain power vs Google ML I was suggesting more simply that the original comment talking about 10x developer wasn't referring to just AlphaGo. Apparently it wasn't clear that I'm limiting my comments strictly to what's observable in the chat forum and not making sweeping statements about humans. Though I suspect self-selection bias might also explain the downvotes. Of course everyone on HN is better than a machine, right? And the brain itself may be. But is the average developer?
- nightski 9y ago
  I didn't down-vote you, but I do disagree. This isn't some desperate attempt to hold on to some sort of artificial superiority to machines. I'd be thrilled if machines could do my job, because honestly half of it is mundane (the other half is communicating and interacting with humans). I'm just being honest about the limitations of current models. I'm also very interested about the impact of generalization. If we make a single model capable of multiple tasks, will that sacrifice the ability to perform on each individual task? Or will it scale?
kingvash 9y ago
Not a huge fan of the article but here it is: https://jacquesmattheij.com/another-way-of-looking-at-lee-sedol-vs-alphago https://jacquesmattheij.com/another-way-of-looking-at-lee-se...