4 ms·

I mean how well do you reckon you can denoise a jpg by hand until its a piece of art? That way of thinking isn’t helpful to understanding AI IMO

by aabhay 1y ago

int_19h 1y ago
In this case it is actually relevant. The ability to draw a pelican on a bicycle correctly depends a great deal on understanding not only what both look like in general, but on the spatial relationships between the various objects and their parts. Models that can draw this kind of thing better also tend to be better at tasks that require understanding of how things go together and interact in 3D space.
- bgwalter 1y ago
  How do we know it's not just a mashup of existing pictures? All generated pelicans on bikes look somewhat cartoonish and use historical or artsy bikes. This is training material from 2015: https://www.behance.net/gallery/29122113/Pelican-on-bikes-with-his-children https://www.behance.net/gallery/29122113/Pelican-on-bikes-wi... There are other such images. Not an image model? How do we know that they don't convert all images to svg and train an LLM on it? How do we know that they do not cheat on this benchmark and route the query to an image model first?
  - jstanley 1y ago
    "it's not impressive because they might have cheated" isn't a great argument.
    - bgwalter 1y ago
      The generated picture is not impressive and the excuse in this subthread was that an svg is created directly without using an image model. I offer alternative explanations why svg creation might not be impressive OR ALTERNATIVELY why they may have faked even a bad result because it is a popular benchmark (faking a perfect result would be too obvious). But since everything is closed source with any number of potential special case hacks, we won't know.
jstanley 1y ago
I didn't intend it as a general-purpose tool for understanding AI, but as an intuition pump for why this problem is hard for LLMs specifically.

I mean how well do you reckon you can denoise a jpg by hand until its a piece of art? That way of thinking isn’t helpful to understanding AI IMO

by aabhay 1y ago

I mean how well do you reckon you can denoise a jpg by hand until its a piece of art? That way of thinking isn’t helpful to understanding AI IMO

int_19h 1y ago

In this case it is actually relevant. The ability to draw a pelican on a bicycle correctly depends a great deal on understanding not only what both look like in general, but on the spatial relationships between the various objects and their parts. Models that can draw this kind of thing better also tend to be better at tasks that require understanding of how things go together and interact in 3D space.

bgwalter 1y ago
How do we know it's not just a mashup of existing pictures? All generated pelicans on bikes look somewhat cartoonish and use historical or artsy bikes. This is training material from 2015: https://www.behance.net/gallery/29122113/Pelican-on-bikes-with-his-children https://www.behance.net/gallery/29122113/Pelican-on-bikes-wi... There are other such images. Not an image model? How do we know that they don't convert all images to svg and train an LLM on it? How do we know that they do not cheat on this benchmark and route the query to an image model first?
- jstanley 1y ago
  "it's not impressive because they might have cheated" isn't a great argument.
  - bgwalter 1y ago
    The generated picture is not impressive and the excuse in this subthread was that an svg is created directly without using an image model. I offer alternative explanations why svg creation might not be impressive OR ALTERNATIVELY why they may have faked even a bad result because it is a popular benchmark (faking a perfect result would be too obvious). But since everything is closed source with any number of potential special case hacks, we won't know.

jstanley 1y ago

I didn't intend it as a general-purpose tool for understanding AI, but as an intuition pump for why this problem is hard for LLMs specifically.