Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

AlphaGo's next move

deepmind.com

Post reply on HN
Print view View original

81–90 of 179 posts

Re: AlphaGo's next move

#81

Post by nopinsight »

temp_12345 wrote: post #75

So, a question bordering on the philosophical : What can be done to prepare for the end of human supremacy, and quite likely human civilization? For instance, as a software developer it feels almost pointless to continue improving at my craft if AI systems will surpass me within 2-4 years (even if the pessimists are right about it taking 5-10 years, that's still an awfully small timeframe). Likewise, it feels a littl…

There are ways to greatly improve chances that AI will be beneficial to humanity rather than otherwise.

Check out: UC Berkeley's Center for Human-Compatible AI, led by Prof. Stuart Russell, a co-author of the field's standard textbook. [1] He just gave a TED talk on the issue [2].

Several other noted researchers in AI are working on the issue as well.

For a short primer: https://futureoflife.org/background/aimyths/

[1] http://www.openphilanthropy.org/focus/global-catastrophic-ri...

[2] https://www.ted.com/talks/stuart_russell_how_ai_might_make_u...

Re: AlphaGo's next move

#82

Post by espadrine »

gwern wrote: post #77
tux3 wrote: post #4

Earlier quoted context omitted.

They did announce that they would release a teaching tool which will show AlphaGo's analysis of Go positions, as well as the paper explaining how to build your own. Not only do you have the principle and the formula behind it, but also a little physics simulator tool! At this point, it is hard to complain.

> At this point, it is hard to complain. Actually, it's very easy to complain. If they released the model, people could generate arbitrarily many self-play games instead of depending on DM to release 50, could create arbitrarily many tools using the model instead of depending on DM to create and maintain a single tool, and could verify the results of training a clone based on even sketchy descriptions of the methods…

> people could generate arbitrarily many self-play games

I have doubts. Their TPU design may be a large factor into making matches at this level within the time limits. And at this point, some implementation details might hook into Google-specific libraries that require the ability to spawn processes in thousands of servers, which past blog posts[0] have hinted at.

[0]: https://deepmind.com/blog/decoupled-neural-networks-using-sy...

Re: AlphaGo's next move

#83

Post by skybrian »

temp_12345 wrote: post #75

So, a question bordering on the philosophical : What can be done to prepare for the end of human supremacy, and quite likely human civilization? For instance, as a software developer it feels almost pointless to continue improving at my craft if AI systems will surpass me within 2-4 years (even if the pessimists are right about it taking 5-10 years, that's still an awfully small timeframe). Likewise, it feels a littl…

I agree that becoming an AI researcher seems daunting, since the field is moving so fast. However, so far we haven't seen any useful applications of machine learning to ordinary computer programming, so this hopelessness seems rather premature.

(Perhaps applying machine learning to code review might be useful, to spot bugs? The problem would be getting good data to train it.)

Re: AlphaGo's next move

#84

Post by skybrian »

lossolo wrote: post #67
adyavanapalli wrote: post #5

Earlier quoted context omitted.

"We plan to publish one final academic paper later this year that will detail the extensive set of improvements we made to the algorithms’ efficiency and potential to be generalised across a broader set of problems." Should be enough, no?

Depends what you look for. Most of ML papers do not disclose weights/models that they used or all details needed to make fully reproducible solution. Doesn't seem like this will change this time.

There are other Go playing programs and they've apparently improved a lot by applying ideas from the original AlphaGo paper. It seems reasonable to assume they will improve more based on ideas from the new paper too, and probably surpass AlphaGo before too long.

(Similarly, Deep Blue was dismantled but chess engines continued to evolve.)

Re: AlphaGo's next move

#85

Post by skybrian »

nopinsight wrote: post #70

One common comment from Go players at all levels up to 9-dan pros is that they don't understand many of the moves. The same will happen as more and more advanced AIs are used in the real world. Yes, we do not completely understand the workings of current advanced neural networks either but the effects are still contained as they are not general enough to cause unintended impact outside their domains. This could have…

They don't fully understand the moves but on the other hand, the live commentary on the games suggests it's not completely mysterious. Good moves still tend to look good to them, in retrospect at least.

The games are apparently very interesting to study.

Re: AlphaGo's next move

#86

Post by artursapek »

temp_12345 wrote: post #75

So, a question bordering on the philosophical : What can be done to prepare for the end of human supremacy, and quite likely human civilization? For instance, as a software developer it feels almost pointless to continue improving at my craft if AI systems will surpass me within 2-4 years (even if the pessimists are right about it taking 5-10 years, that's still an awfully small timeframe). Likewise, it feels a littl…

I've been thinking about this all morning... surely there must be people already working on the following:

- AI that builds an understanding of a large legacy codebase, and is able to diagram & explain it

- using that to refactor convoluted logic and reduce complexity

- using that to train something that can write code from scratch or rewrite an exisiting codebase in a different language

Seems like a billion dollar business, as the world develops more and more large and shitty codebases with high maintenance costs.

Re: AlphaGo's next move

#87

Post by Choco31415 »

wiz21c wrote: post #54
yeukhon wrote: post #51

Earlier quoted context omitted.

Public means API? AlphaGo is very specialized in solving Go game. Google, Amazon and IBM have services for various services like image recongition and speech recongition. Startups like Clarifi also exists in that space. The closest to a generalized AI service would probably be Watson from IBM (but I don't have experiment with it sadly so I am not sure about the usage experience).

Public in the sense of non proprietary, in this case most likely universities. My question was more : Google probably based its tools on existing tools, and those most likely come from universities (research paper, computer infrastructures, etc.). So what are those tools, where are they ?

Google originally used Torch7 as their ML library, but shifted to use TensorFlow in April 2016. TensorFlow is written by Google itself.

AlphaGo itself uses a method that combines Monte Carlo tree search with value and policy NN. All NN used are concurrent NN. The specifics are in a paper by David Silver et al: https://dvc0t0mx8dl84.cloudfront.net/wp-content/uploads/2016...

Re: AlphaGo's next move

#88

Post by cocktailpeanuts »

There was a time not too far back when people used to be considered a "genius" for their ability to memorize things well.

Nowadays nobody thinks of them as geniuses.

Also, people used to be considered geniuses for knowing a lot of things.

Nowadays information is just a Google search away, so knowing a lot doesn't really mean as much as it used to. What matters more nowadays is your ability to learn synthesize the things you know to come up with creative solutions to things.

Basically the "memory" part of human brains have become commoditized without us even realizing.

It's still very early but I do think there have been some subtle but significant step forward in the last couple of years. The most important being: machines are capable of doing certain things better in ways humans can't comprehend easily. I think this is a glimpse into the future where the "creativity" aspect of our brains will become commoditized, also without us realizing.

This doesn't mean machines will take over, just like machines didn't take over the world because they have better memory. But I think this will result in many humans taking advantage of this aspect to exert influence on rest of the humanity.

Re: AlphaGo's next move

#89

Post by skybrian »

yeukhon wrote: post #39
wiz21c wrote: post #37

The gift Google makes to the community (some games AlphaGo,played ) is nothing. The super tricky thing with neural networks is that you can't reverse engineer them. Once the information is coded into the parameters, you can't base anything useful on them. So it's a super good intellectual property protection... Therefore one more nail in the coffin of knowledge sharing as we know it...

Most importanly, it is almost inevitable they must run on a fam of computers basically mean it becomes a service. Can we ever create a robot who can self-learn but with the super brain power locally without having to call a service for an answer?

TPU's are available for rent. Also, apparently AlphaGo became 10x more efficient in a year. Who's to say other Go engines won't continue improving efficiency?

Between software and hardware improvements I think it's likely that we'll see very strong Go engines on desktop computers in a few years.

Re: AlphaGo's next move

#90

Post by gwern »

espadrine wrote: post #82
gwern wrote: post #77

Earlier quoted context omitted.

> At this point, it is hard to complain. Actually, it's very easy to complain. If they released the model, people could generate arbitrarily many self-play games instead of depending on DM to release 50, could create arbitrarily many tools using the model instead of depending on DM to create and maintain a single tool, and could verify the results of training a clone based on even sketchy descriptions of the methods…

> people could generate arbitrarily many self-play games I have doubts. Their TPU design may be a large factor into making matches at this level within the time limits. And at this point, some implementation details might hook into Google-specific libraries that require the ability to spawn processes in thousands of servers, which past blog posts[0] have hinted at. [0]: https://deepmind.com/blog/decoupled-neural-netw…

There might be some hard to release infrastructure code for the MCTS part, certainly, but the model on its own should be a standard TF CNN model and highly competitive (and people can write their own MCTS wrapper, it's not that complex an algorithm). Nothing in the AG paper or statements since has hinted at using anything as exotic as synthetic gradients* and there is no reason to use synthetic gradients in AG. (In RL applications the NNs are generally small because there's so little supervision from the rewards so a large NN would overfit grossly; a NN so large as to require synthetic gradients to be split across GPUs would be simply catastrophicly bad. Plus, the input of a 19x19 board, a few planes of metadata, and other details encapsulating the state is small compared to many applications like image labeling, further reducing the benefits of size. Silver has said AG is now 40 layers but that's not much compared to the 1000-layer Resnet monsters and even those 40 layers are probably going to be thin layers, since it's the depth which provides more serial computation equivalence, not width, making for a model with relatively few parameters overall.)

* I find synthetic gradients super cool and I've been reading DM papers closely for hints of its use anywhere and have been disappointed how the idea doesn't appear to be going anywhere. The only followup so far has been https://arxiv.org/abs/1703.00522 which is more of a dissection and further explanation of the original paper than an extension or application.

Post reply on HN