6 votes

On the Measure of Intelligence

1 comment

  1. onyxleopard
    Link
    I was pointed to this by someone I work with. I think this is an interesting read, and the associated Abstraction and Reasoning Corpus that is proposed is equally interesting. I think theories of...

    I was pointed to this by someone I work with. I think this is an interesting read, and the associated Abstraction and Reasoning Corpus that is proposed is equally interesting. I think theories of intelligence such as Chollet’s and Schmidhuber’s are fascinating and if benchmarks like this are successful, they may push the field of AI closer to achieving GAI sooner rather than later.

    It is very much related to a previous discussion about general artificial intelligence.

    (As an aside, the ARC tests all rely on grids with colored pixels. Being RG colorblind myself, I had to use an on screen color measuring tool to make sure I wasn’t confusing some of the colors before submitting my answers. The paper makes a lot of points about biases in tests, but they failed to consider the bias of using colors as symbols. There are plenty of other ways to signify 10 distinct values without using colors, and even so, they could have chosen a palette that would be colorblind friendly.)

    2 votes