Hacker News Reader

Comment by nohat | Hacker News Reader

Bostrom is in no way in charge of people who happen to agree with him wrt ai risk. For the book he mostly collected and organized a lot of existing thought on ai risk (not that he hasn't made his own novel contributions). That's very valuable, largely because it makes for a good reference point to contextualize discussion on the topic. Unfortunately the critics don't seem to have read it because (in my experience) they repeat the same objections without reference to the existing responses to those objections.

People do sometimes overblow alphago/ dqn playing Atari, but it's not meaningless. These systems (and other deep learning based systems) can truly learn from scratch on a decent variety of environments. One of the most important unknowns is exactly how difficult various cognitive tasks will prove to be for a machine. Each task accomplished is another data point.

View on HN

I wouldn't say that DeepMind learns Atari games "from scratch" any more than Deep Blue learned chess from scratch. It learns to play Atari games because it's a machine designed to learn to play Atari games.

I strongly disagree. You don't seem to be aware of the difference in approach between Deep Blue and DeepMind.

Deep Blue was hand-led directly and specifically to solve the problem of chess: It was provided with a library of opening moves, some sophisticated tactical algorithms relevant to the problem of chess, a library of strategies for chess, and so on. Many actual human masters of chess were consulted, directly or indirectly, to help with developing Deep Blue's approach to the problem.

DeepMind, on the other hand, was created as a "blank slate" with no more hard-wired instruction than "create optimal algorithms to achieve the winning state, given the inputs." Critically, its learning phase is completely self-directed. Essentially, the box is given access to the controls and the video screen content and then sent on its way.

It's instructive to note that this is pretty much exactly how, very generally speaking, evolution and intelligence solve the problem of survival: every organism has controls and a glimpse of "game state" and has to learn (collectively as a species, individually as an organism) to play the game successfully.

> DeepMind, on the other hand, was created as a "blank slate" with no more hard-wired instruction than "create optimal algorithms to achieve the winning state, given the inputs." Critically, its learning phase is completely self-directed. Essentially, the box is given access to the controls and the video screen content and then sent on its way.

Have you seen DeepMind algorithm to be able to say this ? Are there other people outside of Google who have seen the algorithm and can confirm Google's press release?

Maybe you have some legitimate concern about Googles' claim as per their press release and my comment. Who knows, maybe they have some reason to lie about what they did!

But then I wonder why you aren't asking the same question of my parent poster. Has he viewed the DeepMind code, is he qualified to tell us it works the same as chess code? Having made that claim backed on even less evidence than I made mine, I'd say his burden of proof is somewhat greater.

I think there's a heavy dose of press release to what Google is saying. Most people wouldn't call PR puff "lying", but only because standards are low.

I don't think Google has fundamentally different deep-learning technology than everyone else. In fact, TensorFlow indicates that they have the same kind of deep-learning technology as everyone else and they just want to do it more cleanly.

Deep learning is parameter optimization. There are more parameters now, and they optimize more things, but don't get caught up in wild visions of machines designing themselves. Would you consider the bzip2 algorithm to be "self-directed learning"? What's the difference, besides the number of parameters?

The PR people, when they say "blank slate", are discounting all the programming that went into the system because it sounds more impressive that way. This is unfortunate. It has happened in AI for decades. To be a responsible consumer of AI press releases, you need to understand this.

> _I don't think Google has fundamentally different deep-learning technology than everyone else._

That's true, and I never claimed otherwise, but that doesn't help you argue your point - in fact, you just proved yourself wrong. From IBM's press release:

> _Does Deep Blue use artificial intelligence? The short answer is "no." Earlier computer designs that tried to mimic human thinking weren't very good at it. No formula exists for intuition. So Deep Blue's designers have gone "back to the future." Deep Blue relies more on computational power and a simpler search and evaluation function._

I'll summarize for you: Deep Blue and DeepMind, similar names notwithstanding, work in very different ways.

What comparison are you even making here? I know that Deep Blue and Deep Mind are different. There is 25 years (edit: sorry, 20 years) between them! Deep Blue is not deep learning. Did the word "deep", used in two unrelated ways, confuse you?

What I am saying is that I know how deep learning works, actual deep learning of the present, and it does not involve "programming itself".

You are trying to tell me that it must be programming itself, because a press release said so, and press releases would never lie or exaggerate. Based on the current state of AI, this is very improbable. You should focus less on trying to "prove" things with press releases.

I made the comparison to Deep Blue because there is little mystique around it now, and because IBM was even reasonably responsible about avoiding AI hype in their press at the time.

AlphaGo uses TensorFlow.

DeepMind announced the were switching over to use TensorFlow not long ago.

TensorFlow is fully open-source, you can go and read the algorithms on Github. Neural networks aren't particularly hard to understand.

The post you are responding to paraphrases the situation, and neglects to mention goal states/conditions, but is otherwise a fair summary of how neural networks work.

Neural networks, particularly the 'deep' ones, do seem to operate almost like magick. Disbelief is to be expected, but I suggest you take a proper look at them yourself, and see what they can actually do.

If you can grok Python, I highly recommend playing with the Docker image of TensorFlow, which provides a Jupyter notebook (Python in a web interface) where you can easily make use of existing pre-trained models, extend them, and/or build your own models — suitable for application to all kinds of problems / problem domains.

Siraj Raval's videos also cover neural networks in almost every aspect - from coding your own NN from scratch for digit recognition (using just NumPy), through explaining what TensorFlow is, how it works, and how to use it — along with a lot of other machine-learning algorithms, libraries and cloud services.

Many of these topics are just a 5-10 minute video (some with longer live-coding 'deep dive' follow-ups), and many of the code examples are fully functional but just a screen or two of code. https://www.youtube.com/channel/UCWN3xxRkmTPmbKwht9FuE5A

There's also some great new videos on AI / machine learning by Google and Facebook on Youtube that are well worth a watch if the topic interests you.

Does life have a score? A winning state? Exhaustively defined inputs?

I mean, the way you put it makes it sound like you think it very obviously doesn't. But are you willing to grant that maybe there's some very complicated optimization function for life satisfaction/personal fulfillment/eudamonia/whatever the hell you want to call it? It doesn't have to be practically computable to be interesting, you merely have to be able to do your best to approximate it.

If you deny the existence of any such thing, isn't that tantamount to saying that given the choice between two courses of action, there's never a reason to pick one over the other?

I mean, I feel like as humans, we're trying to follow fuzzy approximations of that function all our lives, whenever we're earning money, or spending time with people we love, or whatever, we're doing it because we think it will bring us happiness or life satisfaction or something.

sometimes there isn't a correct choice... both could be good, both could be bad. The Road Not Taken: http://www.bartleby.com/119/1.html

Absolutely! The goal of the game is to reproduce. To be more specific, it's reproduction at the individual level and survival at the species level.

The genome is a kind of algorithm for hardware to implement a strategy to win at this game in a highly competitive environment.

The Atari AI's learning phase is a specifically designed deep neural network. The network did not design itself. It was designed by human programmers.

There are probably numerous variants of it that did not learn to play Atari games, and therefore were not announced.

The human brain didn't design itself either - it's the product of millions of years of evolution via natural selection. But this fact is irrelevant to the topic of whether the approach of DeepMind is fundamentally different from the approach of Deep Blue.

To help you appreciate the difference, try answering this question: Were experts players of Pong, Space Invaders or any other video games consulted to contribute strategies for winning those particular games? Was a redesign required to play a new, different game?

If not, you'll need to tell me where the knowledge of how to win at those games came from. I hope you'll appreciate that the ability to autonomously obtain that knowledge by trial and error rather than design and programming constitutes the fundamental difference.

Any actual chess knowledge in Deep Blue was an optimization. The fundamental algorithm is alpha-beta search, which can learn to play many, many games of strategy. Not Go, as we know now, but the exceptions are few. This is an algorithm that is very adaptable and very good at learning to play games based on a heuristic for whether you are winning.

The DeepMind implementation learns how to play Atari games. You could even say that it's learning how to play one game, where your opponent chooses the game cartridge and your goal is to increment the score counter.

You would absolutely need to redesign DeepMind to play Super Mario Bros.

When you give DeepMind credit for designing itself, you are glossing over the fact that it was programmed by people, most likely for long hours.

Reply on news.ycombinator.com