from Hacker News

Why AI can ace medical board exams but can't make real art

by kurinikku on 1/16/25, 2:11 AM with 2 comments

• by onecommentman on 1/16/25, 5:41 AM

  I’ve become less impressed with AI performing well on standardized tests. I’m not tracking AI daily so this might be a known insight, but nearly the entire space of a given discipline’s standardized test questions is enumerable by today’s computers, or very close. That’s my guess. These standardized test questions are generally of a very structured format, even Math Olympiad. “Why that’s still billions and billions of possible questions” thinks the average person. And everyone on this forum knows that this is easily manageable with today’s technology.

  So the computer can train on almost all possible standardized test questions, effectively memorizing their answers in a very compressed format with a very interesting adaptive compression algorithm. So for a AI, standardized tests are open book exams. But those questions are designed to challenge human students in closed book exams under human time constraints. So, who cares?

  Where is my reasoning faulty?

• by tolerance on 1/16/25, 2:30 AM

  Art is inherent and exclusive to human beings, and exams irrespective of the subject are gauges of rote ability.

  We don’t need graphs and charts to discern this.

  Robot art is what you get if Hitler was born in 2002 and learned how to paint watching Snapchat stories.

  > "[Prosaic], utterly devoid of rhythm, color, feeling, or spiritual imagination. They are architect's sketches: painful and precise draftsmanship; nothing more.”

  https://en.m.wikipedia.org/wiki/Paintings_by_Adolf_Hitler#Cr...