Futurist Ray Kurzweil Pulls Out All the Stops (and Pills) to Live to Witness the Singularity • Hacker News Board

Re: Futurist Ray Kurzweil Pulls Out All the Stops (and Pills) to Live to Witness the Singularity

#91

Post by bumbledraven » Sun, Mar 30, 2008, 7:04 AM UTC

gruseom wrote: post #90

bumbledraven wrote: post #89

Earlier quoted context omitted.

[I]t depends on whether or not there's an algorithm for everything. We already know that there is not an algorithm for everything. Indeed, Turing himself showed that there is not an algorithm for solving the halting problem " rel="nofollow">http://en.wikipedia.org/wiki/Halting_problem> . But that has no bearing on whether or not there exists an algorithm that would someday allow a computer to pass the Turing Test.

Replace that phrase with algorithm for everything that's intrinsic to being human . Or, if you prefer, algorithm for everything necessary to simulate being human sufficiently well to consistently fool humans via low-bandwidth TTY . My point remains the same: it's a belief either way. And throwing processing power at it will only yield the predicted effect in one of those cases. Edit: the existence of undecidability r…

[I]t depends on whether or not there's an algorithm for everything necessary to simulate being human sufficiently well to consistently fool humans via low-bandwidth TTY.

But clearly such an algorithm does exist, so this objection fails.

Its existence follows from these facts:

(1) only a fixed amount of data in the form of questions can flow through the low-bandwidth TTY,

(2) only a fixed amount of data in the form of responses can be sent back back, and

(3) there exists an algorithm for any function with a bounded input and output. (To see this, note that the algorithm could consist of a lookup table containing a correct output for each of the finite number of inputs it has to deal with.)

I concede that it is another matter entirely whether or not we will ever discover such an algorithm and implement it on a sufficiently powerful computer. But the fact that an algorithm for passing the Test must exist takes the discussion out of the realm of physical and/or philosophical possibility and into the realm of technology and engineering.

Re: Futurist Ray Kurzweil Pulls Out All the Stops (and Pills) to Live to Witness the Singularity

#92

Post by gruseom » Sun, Mar 30, 2008, 6:55 PM UTC

bumbledraven wrote: post #91

gruseom wrote: post #90

Earlier quoted context omitted.

Replace that phrase with algorithm for everything that's intrinsic to being human . Or, if you prefer, algorithm for everything necessary to simulate being human sufficiently well to consistently fool humans via low-bandwidth TTY . My point remains the same: it's a belief either way. And throwing processing power at it will only yield the predicted effect in one of those cases. Edit: the existence of undecidability r…

[I]t depends on whether or not there's an algorithm for everything necessary to simulate being human sufficiently well to consistently fool humans via low-bandwidth TTY. But clearly such an algorithm does exist, so this objection fails. Its existence follows from these facts: (1) only a fixed amount of data in the form of questions can flow through the low-bandwidth TTY, (2) only a fixed amount of data in the form of…

Uh oh, you got technical on me!

It's been a long time since I've studied this stuff, but for what it's worth, your argument feels incorrect to me. It might not be possible to address that effectively here, but let me try.

Let's call a particular run of the Turing test a "Turing dialog". Say a Turing dialog D consists of alternating statements C1,H1,...,Cn,Hn where C stands for "computer" and H for "human". Say a "successful Turing dialog" is one that succeeds in fooling the human participant - that is, Hn is something like "I think you're human".

You're right that for any Turing dialog D, D is finite and therefore an algorithm exists to reproduce it (a simple lookup table will do). But such an algorithm is only good for one specific D. To build it, you'd have to know all of D in advance.

To pass the Turing test, we need more. We need an algorithm, which, given any valid prefix of a Turing dialog, i.e. any sequence C1,H1,...,Ck,Hk, knows how to produce Ck+1 in such a way that the dialog will end successfully. That's not the same thing, is it?

Re: Futurist Ray Kurzweil Pulls Out All the Stops (and Pills) to Live to Witness the Singularity

#93

Post by bumbledraven » Sun, Mar 30, 2008, 7:24 PM UTC

gruseom wrote: post #92

bumbledraven wrote: post #91

Earlier quoted context omitted.

[I]t depends on whether or not there's an algorithm for everything necessary to simulate being human sufficiently well to consistently fool humans via low-bandwidth TTY. But clearly such an algorithm does exist, so this objection fails. Its existence follows from these facts: (1) only a fixed amount of data in the form of questions can flow through the low-bandwidth TTY, (2) only a fixed amount of data in the form of…

Uh oh, you got technical on me! It's been a long time since I've studied this stuff, but for what it's worth, your argument feels incorrect to me. It might not be possible to address that effectively here, but let me try. Let's call a particular run of the Turing test a "Turing dialog". Say a Turing dialog D consists of alternating statements C1,H1,...,Cn,Hn where C stands for "computer" and H for "human". Say a "suc…

We need an algorithm, which, given any valid prefix of a Turing dialog, i.e. any sequence C1,H1,...,Ck,Hk, knows how to produce Ck+1 in such a way that the dialog will end successfully.

Exactly, but since each Turing dialog is of bounded length (i.e. no more than a megabyte in size, depending on the bandwidth of the TTY and the predetermined maximum length of the test), the set of all Turing dialogs is finite and is therefore subject to the dictionary algorithm approach.

It's rather like building out the game tree of optimal play for White in chess. The tree for the Turing Test "game" would have to include all the questions that could be asked at each point in the tree, but you only need one possible correct ("Turing-test passing") response at each node.

Re: Futurist Ray Kurzweil Pulls Out All the Stops (and Pills) to Live to Witness the Singularity

#94

Post by Novash » Tue, Apr 01, 2008, 12:03 AM UTC

aswanson wrote: post #78

Novash wrote: post #74

Earlier quoted context omitted.

Solved? I think this qualifies within his predictions. http://www.technologyreview.com/blog/editors/22037/

I'm guessing that given the difficulty of characterizing voiced speech is still inaccurate, the more difficult task of translating neural impulses to speech will result in even further degradation of performance. Unsolved: At the moment, the device has a limited vocabulary: 150 words and phrases. A long way to go.

They still have an year and a half, which means twice the number of transistors on their next hardware. Give them the benefit of doubt.

Re: Futurist Ray Kurzweil Pulls Out All the Stops (and Pills) to Live to Witness the Singularity

#95

Post by aswanson » Wed, Apr 02, 2008, 1:25 AM UTC

Novash wrote: post #94

aswanson wrote: post #78

Earlier quoted context omitted.

I'm guessing that given the difficulty of characterizing voiced speech is still inaccurate, the more difficult task of translating neural impulses to speech will result in even further degradation of performance. Unsolved: At the moment, the device has a limited vocabulary: 150 words and phrases. A long way to go.

They still have an year and a half, which means twice the number of transistors on their next hardware. Give them the benefit of doubt.

No, I won't. The problem is not transistors; the problem is in the lack of understanding how speech recognition works. It is very difficult to get right and it is highly unlikely that those problems will be solved completely within the next 18 months.

They could double the transistor count now by running the algorithm on a better DSP or Cell processor.