8 ms·
Google’s AlphaGo AI defeats team of five leading Go players
- arcanus 9y agoAlgorithms 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.
- nightski 9y agoWith 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.
- blahbwah777 9y agoIMO, 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 agoI'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 agoThe brain is efficient but, let's be honest, not really productive, especially not 100% of the time.
- obstinate 9y agoEven if it is, that's a questionably useful observation, because computers can consume and dissipate much more energy than the human brain.
- narrator 9y agoLuckily 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 agoYou say that now, but then it installs a Dyson sphere. Very dark indeed.
- philipkglass 9y agoThis 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 agoSure, 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 agoI 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 agoNot 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...
- arcanus 9y agoGreat 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!
- carbocation 9y agoIs 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.
- echevil 9y agoAt this point no one expect alphago would lose. So just have fun
- IanCal 9y agoI'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?
- Jenya_ 9y agoThere is already a system in Go which allows to measure the difference between unequal opponents - handicap stones.
- Jach 9y agoAnd 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.
- Ajedi32 9y agoYeah, 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.
- mackncheesiest 9y agoTo 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 :)
- Florin_Andrei 9y agoWell, what do you think your brain is? :)
- platz 9y agoa series of interconnecting drives, shafts, and levers?
- EternalData 9y agoI thought this was supposed to be a few years away -- i.e beating Lee Sidol was impressive, but Lee wasn't nearly the top player in the world? Wow, this must be what it feels like to be living on a parabolic curve...
- daveguy 9y agoLee Sedol was near the top (#7). He wasn't the very top, that was Ke Jie. And now Ke Jie has been beaten 2/3, maybe 3/3 times. 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/ https://www.goratings.org/en/
- stupidcar 9y agoIronic that even though Sedol is not the strongest human player, he may end up being the last human in history to beat the strongest AI.
- thomasahle 9y ago> 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).
- Deimorz 9y agoAccording to some information that the AlphaGo team released earlier in the week, the current version of AlphaGo can give a three-stone handicap to the version that played Lee Sedol: http://www.usgo.org/news/2017/05/new-version-of-alphago-self-trained-and-much-more-efficient/ http://www.usgo.org/news/2017/05/new-version-of-alphago-self.... 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.
- cjbprime 9y ago
- 31reasons 9y agoThis is going to make all the Go players out of job pretty soon. :)
- pdpi 9y agoNo more than Deep Blue put pro chess players out of a job.
- Jach 9y agoIt 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?
- cyborgx7 9y agoCars have been faster than people for ages, but marathons and sprints are still popular sports.
- jayramone 9y agoMachine vs human is never fair. That is like Usain Bolt vs a modern car.
- chr15 9y agoThis 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.
- sprafa 9y agoI'd say human intelligence hasn't reached a ceiling, only that it doesn't grow as quickly as machine intelligence can.
- tokai 9y agoHave 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.
- foobar__ 9y agoWhere did you find this information for chess? It sounds unintuitive that a superior player would benefit from a weak player's help.
- tokai 9y agoI 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] [0]https://www.bloomreach.com/en/resources/blogs/2014/12/centaur-chess-brings-best-humans-machines.html; https://www.bloomreach.com/en/resources/blogs/2014/12/centau... http://www.huffingtonpost.com/mike-cassidy/centaur-chess-shows-power_b_6383606.html; http://www.huffingtonpost.com/mike-cassidy/centaur-chess-sho... http://www.glassbeans.com/blog/why-centaurs-will-dominate-the-future/ http://www.glassbeans.com/blog/why-centaurs-will-dominate-th... [1]https://chess.stackexchange.com/questions/15772/how-do-the-best-centaurs-compare-to-the-best-computers https://chess.stackexchange.com/questions/15772/how-do-the-b...
- foobar__ 9y agoI 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.
- farhanhubble 9y agoAn interesting thing to observe in the long run would be if humans playing against AI are able to improve their moves and beat the AI.