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.
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.
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.)
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?
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?
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!
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?
And it's interesting to note that according to DeepMind, the Master version of AlphaGo was 3 stones stronger than the version that played Lee Sedol. I don't know if the latest one playing Ke Jie is even stronger than Master or by how much, but the difference in Ke Jie and Lee Sedol isn't 3 stones.
Yeah, that was a question I had while watching the stream too. Is a team of Go experts working together typically considered stronger than a single, highly skilled player playing alone?
Also worth noting that due to the time constraints of this particular match, the human players didn't really have a lot of time to discuss and debate every move they chose, which could have had a negative impact on the potential advantages of working as a team.
If you consider that the team is just a group of experts discussing an ongoing game and variations that may be better/worse, it seems pretty common. I'd imagine that multiple experts could perform better given sufficient time. With only a few hours, most of the time seems like it'd be lost in inefficient human IPC.
Pair Go matches are trendy these days from what I read. It's a good training process for stronger player teaching the relatively weaker player in the pair in real matches.
Team Go playing is probably a different story I agree.
I always said in Pair Go , that 1 bad idea is better than 2 good ones.
But provided discussion, you can be a lot more thorough in consensus. Professional analyze and study games in groups, and some thing come out of that.
I'd say overall the result is a game with no blunders , but no edge. Since you are sharing the blame of a loss, you are not as focused on winning as you are on not losing.
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.)
Playing go cooperatively is very different from chess or most other turn-based tactical games in that combinatorics make the gamespace astronomically large. Most pro players understand what impact this has on gameplay and how to adapt their risk management strategies to the game. Way back in the day, popular man v cpu strategies took into account the limitations of monte carlo simulations and would intentionally make moves that change the game state as much as possible to exploit how hard the computer would need to work in order to begin finding optimal moves.
I think alpha go has advanced to the point where its unlikely human players can reliably defeat it, but I think there are still opportunities for better algorithmic players to defeat it.
Anyways, certain moves can cut off game states by the quadrillions, so it's often pretty intuitive what an optimal response to a move is, given certain context of the board and sometimes the player. In that respect, I'm curious how pro level go is going to change after these alpha go games are studied, because it has a very peculiar, decidedly calculatorish style of play, but it's obviously very effective regardless. Go has prospered for so long because there's so much room to express yourself in a move, pro players really play with their whole soul, but computers are just taking advantage of the pure mathematical angles of the game
It's fascinating stuff. I really want to see the alpha go team write a starcraft ai or something like that.
My thoughts were on this line also... the same reason the crowd was shocked when Lee Sedol made an unexpected move in game 4. [0]
Is there any research done to prove/disprove that groupthink is inherently worse compared to a single genius? (less risky behavior, so less risk/less reward??) Are we training AIs that will be safe, not bold independent thinkers?
> Gu Li was talking about previous Team matches, and how once a team of Shi Yue, Zhou Ruiyang and Chen yaoye kept on arguing about what to do, and couldn't reach an agreement...
> Meanwhile, the other team had Kang Dongyun, Park Junghwan and Choi Cheolhan. Choi Cheolhan occasionally looked at the variations to make sure there are no silly mistakes, Kang Dongyun's job was to buy lunch for everyone, and Park Junghwan played the game.
> this is an improvement, but not an exponential one.
I wonder how you define exponential here. If the old version had a probability 20% of losing against Lee Sedol, and the new one has 5%, then one might call it exponential. Something like losing prob = 2^(2012-current year).
I don't know if that qualifies as "exponential", but that's a massive difference in skill at the game. If it's accurate at all, I think there's basically zero chance that a human will ever beat AlphaGo again.
Terrifying in some ways, but I think it's also inspiring. It's pretty amazing to have some evidence that there's still that much "room" left at the top of the game for players to continue improving.
And the more interesting note - it gained a three stone handicap of skill with 1 year of practice.
If they were to continue at it should we expect a 3 stone improvement in skill every year? Maybe they made some major leaps this time and future changes aren't as large. So a 3 stone improvement every two years.
It doesn't take long until the Go being played is fundamentally a different game than what we play. Even if it's on the same board with same rules.
What does it look like when Go still isn't solved, but AI is 20 stones better than the best humans? At that point we can no longer even tangibly measure its brilliance. Only ask it to compare its brilliance to other brilliant AIs. What will that even look like through out eyes? What will like look like when an AI is 20 stones better than us at war games?
I assume next step for AlphaGo is a real time strategy game like Starcraft, or some other game with incomplete information. 3-5 years from now when mastered it'll be hard to persuade people not to use it in the real world.
AlphaGo had a fundamentally new architecture (nobody had successfully applied convolutional neural networks to Go before, although of course the headline underestimates the expertise and experimentation that's needed to get the details right).
So it's not surprising that at the first public announcement they hadn't yet pushed the technology as far as it would go.
It should be interesting to see how a prevalence of stronger-than-human go bots will influence the next generation of players. Magnus Carlsen relies on Chess bots extensively for analysis and preparation (though not really as opponents since they're too good), how much stronger than Ke Jie is now will be the Go player of the future who comes of age using such tools? And for the existing pros already, will we soon see a large divide in those who embrace using such tools (and studying their games with each other / good humans as has already been done with the Master series) and those who dislike them?
This is pretty amazing. It shows there is a level of intelligence that exists above what these 5 can achieve collectively. Scientists have effectively created life when something can make decisions for itself and outsmart humans.
Human intelligence has reached a ceiling and machines will get smarter and smarter. In hundreds or thousands of years, AI will be smart enough to control us. This Go game will manifest itself into reality - it will be a strategic game of humans vs machines.
Have they done a human and AlphaGo vs AlphaGo yet? As I understand it centaur teams of human and computer in chess are superior to computers or humans playing alone. It would be interesting to see if this holds with go as well.
I can't give you the source I got it from, as it is something I have read multiple times over the years from different sources. A quick search turns up a lot of articles on the subject, with the claim that centaur teams are better.[0]
I cannot follow you on the unintuitive part. If you have a chess program playing against itself, as long as it isn't capable of playing a perfect game every time, the program being supervised by a human should better, as the human will have a non-zero chance of spotting a mistake (or a better move) with enough plays. With low of the risk of mistakes from the weaker human, as she can confer with the judgement of the program. If the risk of human error is lower than the chance of human insight, the centaur team should have the edge on the lone computer.
After checking up on the subject again, it seems like the effect is no longer strong enough to topple the best chess computers though.[1]
I find it unintuitive when the strength difference between the two players is too high, and I think matter-of-fact statements like in the stackoverflow answer ("Of course, a human assisting a computer player will be stronger than the same computer player alone") need to be questioned.
Imagine a beginner receiving suggested moves from a grandmaster. I can't see how this beginner could be stronger than the grandmaster alone, and they would almost certainly be weaker if they override just one of the grandmaster's suggestions.
From what I read, chess AIs are this far above humans.
There are centaur championships in chess, for example ICCF. Mainly, people try do strategy and choose directions (which engines are not that good at) , while computer does the calculations.
56 comments
[ 4.7 ms ] story [ 112 ms ] threadI'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.
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.
Sadly I have no recollection of where I've read this.
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.)
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?
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?
https://jacquesmattheij.com/another-way-of-looking-at-lee-se...
And of course, it took 100+ top PhDs in ML to create DeepMind. It certainly required many humans!
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?
Also worth noting that due to the time constraints of this particular match, the human players didn't really have a lot of time to discuss and debate every move they chose, which could have had a negative impact on the potential advantages of working as a team.
Team Go playing is probably a different story I agree.
I always said in Pair Go , that 1 bad idea is better than 2 good ones.
But provided discussion, you can be a lot more thorough in consensus. Professional analyze and study games in groups, and some thing come out of that.
I'd say overall the result is a game with no blunders , but no edge. Since you are sharing the blame of a loss, you are not as focused on winning as you are on not losing.
FWIW, the five people team played with Ke Joe before the game and won (again, according to Gu Li.)
I think alpha go has advanced to the point where its unlikely human players can reliably defeat it, but I think there are still opportunities for better algorithmic players to defeat it.
Anyways, certain moves can cut off game states by the quadrillions, so it's often pretty intuitive what an optimal response to a move is, given certain context of the board and sometimes the player. In that respect, I'm curious how pro level go is going to change after these alpha go games are studied, because it has a very peculiar, decidedly calculatorish style of play, but it's obviously very effective regardless. Go has prospered for so long because there's so much room to express yourself in a move, pro players really play with their whole soul, but computers are just taking advantage of the pure mathematical angles of the game
It's fascinating stuff. I really want to see the alpha go team write a starcraft ai or something like that.
Is there any research done to prove/disprove that groupthink is inherently worse compared to a single genius? (less risky behavior, so less risk/less reward??) Are we training AIs that will be safe, not bold independent thinkers?
0: https://www.wired.com/2016/03/two-moves-alphago-lee-sedol-re...
https://youtu.be/V-_Cu6Hwp5U?t=4h26m30s
There are other points in the commentary where they discuss some of the team's strategy, as well as their time management.
> Gu Li was talking about previous Team matches, and how once a team of Shi Yue, Zhou Ruiyang and Chen yaoye kept on arguing about what to do, and couldn't reach an agreement...
> Meanwhile, the other team had Kang Dongyun, Park Junghwan and Choi Cheolhan. Choi Cheolhan occasionally looked at the variations to make sure there are no silly mistakes, Kang Dongyun's job was to buy lunch for everyone, and Park Junghwan played the game.
> The Korean team ended up winning.
Wow, this must be what it feels like to be living on a parabolic curve...
Sedol was beaten soundly -- 4/5 games. So this is an improvement, but not an exponential one.
Go is definitely conquered like chess was 1-2 decades ago.
https://www.goratings.org/en/
I wonder how you define exponential here. If the old version had a probability 20% of losing against Lee Sedol, and the new one has 5%, then one might call it exponential. Something like losing prob = 2^(2012-current year).
I don't know if that qualifies as "exponential", but that's a massive difference in skill at the game. If it's accurate at all, I think there's basically zero chance that a human will ever beat AlphaGo again.
If they were to continue at it should we expect a 3 stone improvement in skill every year? Maybe they made some major leaps this time and future changes aren't as large. So a 3 stone improvement every two years.
It doesn't take long until the Go being played is fundamentally a different game than what we play. Even if it's on the same board with same rules.
What does it look like when Go still isn't solved, but AI is 20 stones better than the best humans? At that point we can no longer even tangibly measure its brilliance. Only ask it to compare its brilliance to other brilliant AIs. What will that even look like through out eyes? What will like look like when an AI is 20 stones better than us at war games?
I assume next step for AlphaGo is a real time strategy game like Starcraft, or some other game with incomplete information. 3-5 years from now when mastered it'll be hard to persuade people not to use it in the real world.
So it's not surprising that at the first public announcement they hadn't yet pushed the technology as far as it would go.
Human intelligence has reached a ceiling and machines will get smarter and smarter. In hundreds or thousands of years, AI will be smart enough to control us. This Go game will manifest itself into reality - it will be a strategic game of humans vs machines.
I cannot follow you on the unintuitive part. If you have a chess program playing against itself, as long as it isn't capable of playing a perfect game every time, the program being supervised by a human should better, as the human will have a non-zero chance of spotting a mistake (or a better move) with enough plays. With low of the risk of mistakes from the weaker human, as she can confer with the judgement of the program. If the risk of human error is lower than the chance of human insight, the centaur team should have the edge on the lone computer.
After checking up on the subject again, it seems like the effect is no longer strong enough to topple the best chess computers though.[1]
[0]https://www.bloomreach.com/en/resources/blogs/2014/12/centau... http://www.huffingtonpost.com/mike-cassidy/centaur-chess-sho... http://www.glassbeans.com/blog/why-centaurs-will-dominate-th...
[1]https://chess.stackexchange.com/questions/15772/how-do-the-b...
Imagine a beginner receiving suggested moves from a grandmaster. I can't see how this beginner could be stronger than the grandmaster alone, and they would almost certainly be weaker if they override just one of the grandmaster's suggestions.
From what I read, chess AIs are this far above humans.