--- Benjamin Goertzel <[EMAIL PROTECTED]> wrote:
> And yet, I can easily imagine a narrow-AI approach to creating new funny
> videos, e.g.
> 
> 1 -- make an image-processing system that can map out objects and
> relationships in a video [e.g. fall_in(dog, toilet) ... fall_in(pig, lake)
> ... ]
> 
> 2 -- run a machine learning algorithm on the relation-sets extracted from
> videos, with the goal of learning rules distinguishing funny from unfunny
> videos
> 
> 3 -- write a program to generate animated videos from sets of relationships
> [e.g. to go from fall_in(dog, toilet) to an animation of a dog falling in a
> toilet]
> 
> 4 -- use the rules learned in Step 2 to generate new videos [possible e.g.
> if the rules embody a probability distro; then Step 4 is Bayesian instance
> generation...]
> 
> A system like this may be constructible without any kind of deep AGI or
> insight into humor.  It might not work perfectly, but humans' sense of humor
> is not perfect either.  Yet it would lack any real insight or originality.
> And it would only do this one application, without ability to generalize.
> 
> We have not yet plumbed the full depths of narrow AI, I feel.  Nowhere near.

But what if I like political humor?

And recognizing video of a dog falling in a toilet doesn't seem like an easy
problem to me.

But let's simplify the problem.  Is it possible to write a text-only joke
detector?  Exactly what makes a joke funny?

Suppose we take thousands of jokes with various ratings for degrees of
funniness and feed it to a machine learning algorithm.  Ignore the fact for
now that a joke isn't funny the second time you hear it.  What features are
you looking for?  What features make a joke funny?  Can you even tell me why a
particular joke is funny or not funny?  What qualities of "bad AGI poetry" did
you use to pick the ones to post on your blog, and could you have written a
program to do it?

If this is narrow AI, why hasn't somebody done this 20 years ago?


-- Matt Mahoney, [EMAIL PROTECTED]

-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=e9e40a7e

Reply via email to