Hacker News Reader

Comment by ForHackernews | Hacker News Reader

They made this bet in 2002, so we're almost halfway to 2029. Does anyone (other than Kurzweil) seriously think a Turning Test-passing machine is just over the horizon?

(And no, contrived scenarios with computers pretending to be foreign children don't count[0])

[0] http://blogs.wsj.com/digits/2014/06/10/did-eugene-goostman-p...

View on HN

But Kurzweil is banking on the exponential growth of technology. According to his reasoning (and using Moore's Law of 2x performance every 18 months, as a guideline) we are less than 1% of the way towards the goal (assuming 100% is passing the Turing test).

Now I don't have any idea if he will be correct, or even what a machine that will pass the Turing test will be like. But if you accept the premise that technological advancement is on an exponential curve, half way in terms of progress will be much closer to 2029 than half way in time.

(Note I am not saying that technology is exponential or that it is following Moore's law or that it will be exponential forever or that 2x every 18 months is fact, its simply an extension of Kurzweil reasoning to its logical conclusion.)

Exponential improvements aren't very helpful if you're working on exponentially hard problems.

Exponential growth of technology hasn't solved machine intelligence (by direct simulation) for essentially the same reason why it hasn't solved numerical weather prediction or made quantum computers pointless.

To expand on weather prediction: you need an order of magnitude improvement in computer speed to make weather forecasts with the same quality one day earlier. So exponential growth of computing power means linear growth in terms of subjective benefit.

To be fair, we don't understand the difficulty of modeling intelligence anywhere near as well as we understand the difficulty of modeling the weather. But, given the limits of our current abilities to simulate the brain, it seems reasonable to guess that similar principles could hold.

What's exponentially hard about it?

Talking like a human is a bounded problem, it has no margin for growing (or shrinking). Any kind of growth in capacity will help against it.

Fair enough, but that's like saying "intergalactic travel is a bounded problem"

Well, it is. You can't claim it grows in any specific way, because it does not grow (at least at the Local Group).

You can not claim that exponential growth on space travel tech isn't enough to get intergalactic travel. Any non-declining rate of progress will get us there eventually. (Whatever "exponential growth" means on this context, for computation it's very well defined. Also, whatever realist is on non-declining rate of progress, the rate for computation is clearly increasing.)

>Exponential improvements aren't very helpful if you're working on exponentially hard problems.

Sure they are! They reduce it to a problem that's only polynomially hard! :-p

>Exponential growth of technology hasn't solved machine intelligence (by direct simulation) for essentially the same reason why it hasn't solved numerical weather prediction or made quantum computers pointless.

Okay, but that's a rather high bar to judge it against. Humanity has solved many lesser machine intelligence problems, like web search, voice recognition, routing, and recommendation. Though I agree the relevant exponential growth was in the power of the algorithms, not so much the hardware.

I don't see why "weather forecasting" is similar to "brain simulation" in difficulty. The comparison would be more valid if the problem were "brain forecasting", that is to build an AI that would behave exactly the same as a particular human, but that's an even harsher measure of intelligence than the turing test that I doubt anyone would take seriously. I'm not saying the brain's easy to simulate, it's just the analogy here seems shaky. Where is the source of exponential difficulty?

If you have no idea what you are doing then doing it exponentially faster won't help much.

I think there's an excellent chance that deep learning research will lead to a machine that can pass the Turing test in the not too distant future. I can't say if it will be within exactly 14 years or not, but if you've been following the latest developments in deep learning, the path to get there is much more clear today than it was even five years ago.

I think you are greatly overestimating what deep learning can do. In the 90s, we could recognise digits accurately. Now we can do the same with traffic signs even in bad weather etc. That is exactly the kind of progress that we have made in 14 years. And let's not forget this is a manually tuned algorithm for a particular problem.

It's great progress, but it is also a far cry from what humans can do and there is no clear path at all to get there - currently.

The recent advancements using RNN is much more than sign recognition. It's about language generation, control and reinforcement learning, and attention.

I assume that deep learning will help with coming up with suitable replies. But will it also helping with keeping track of state, ie. making sure that the computer's answers are consistent? It seems to me that this will be the hardest part about passing the Turing test.

Re keeping track of state: https://www.facebook.com/video.php?v=10153098860532200

Could you give some info on where to keep up with the latest on deep learning for the quasi-layman (I've got a bachelors in CS)?

From what I've seen arXiv has a lot of papers that are new and related. It might be a good place to start.

agreed - there are current approaches that are becoming feasible as infrastructure is opened up and affordable to the general public. It's a game changer but a lot of people misinterpret the meaning and implications it seems like. A good question to ask is how this advance would affect you directly.

Given the specific 'Turing Test' procedure described in the bet, I think there's an excellent chance Kurzweil will win. (If the IM sessions were 5 minutes with each judge rather than 2 hours, he might win today.)

2029 is a long way still and I would be surprised if the Turing Test wasn't passed. The past few years have had massive progress in AI. A few years ago machine vision was an impossible problem believed to require strong AI, and now we are exceeding humans.

Massive progress is being made in natural language understanding too. A number of papers have just come out on using deep NNs to do question answering or conversation.

> now we are exceeding humans.

I mean, for some particular, limited version of "exceeding", yes: http://www.wired.com/2015/01/simple-pictures-state-art-ai-st...

Those are images incredibly optimized to fool and exploit those neural networks. However on the actual image recognition tasks they were designed for they beat humans.

Anyway this paper has a method for addressing this problem: http://arxiv.org/abs/1412.6572

That's exactly my point. Computer vision is a "solved" problem within certain specific domains for which a particular algorithm has been trained/tuned ("the actual image recognition tasks they were designed for"). They are nowhere near as good as humans at the general-looking-at-any-random-things task yet.

Similarly, I imagine we will get very good domain-specific "AI" in the next ten years: some sort of amped-up Siri that can, say, answer questions about hotels in a travel destination and book one for you.

I do not think we will get a chatbot that can hold a conversation on any conceivable topic at a level that you would confuse for a human.

Imagenet is a pretty open task. The images are taken randomly from the internet with real world clutter and noise and other issues. There are 1,000 classes which is pretty large compared to previous datasets.

When it started getting attention a few years ago algorithms sucked at it. No one would have believed in a few years we could beat humans on it. It seems like a case of moving goal posts, where every time a milestone is reached, it's disregarded.

> It seems like a case of moving goal posts, where every time a milestone is reached, it's disregarded.

I think it's sort of a case of AI researchers tried moving the goalposts closer, and now we're moving them back. Back in the 60s, a bunch of people thought a general-purpose artificial intelligence was a realistic goal in the near future.

Then it turned out that problem was hard, really hard. AI researchers instead defined-down the problem. Instead of building AI, they worked on building "chess-playing AI" or "traffic-optimizing AI". These are useful and interesting in their own right, but they're not really what people mean when they talk about "intelligence".

Over time, this meant that the entire field of "AI" became discredited, which is why nobody talks about AI anymore, they talk about "machine learning", or "deep learning", etc. If somebody says they're going to make a human-level general-purpose AI, they're no longer taken seriously.

So what metric are you using to measure progress in AI? There are plenty of benchmarks on very general tasks like image recognition or language modeling, etc. There is no room for bias or moving goal posts on a benchmark.

All I'm saying is that AI is rapidly advancing. I'm not claiming its human level yet, and I don't really care what people in the 50s predicted.

Do you happen to have any links to any of those papers? I would love to read them.

I was thinking mostly of this paper:

http://arxiv.org/abs/1503.02364

Here is some other relevant work of recent:

http://arxiv.org/abs/1502.05698

http://arxiv.org/abs/1410.3916

https://www.reddit.com/r/thisisthewayitwillbe/comments/2zi8i...

https://www.reddit.com/r/thisisthewayitwillbe/comments/25a4r...

https://www.reddit.com/r/thisisthewayitwillbe/comments/307pt...

http://www.noahlab.com.hk/topics/ShortTextConversation

http://allenai.org/content/publications/hixon_naacl_2015.pdf

Thanks!

Reply on news.ycombinator.com