A few days ago, I got a job as a remote AI trainer. Since I still work with the same company, I won't disclose its identity. They gave me a screening test and said my mathematical ability was superior. Thus, I was assigned to train an AI math solver. I wasn't sure what to expect, so I started out with some high school and college level algebra problems. My instructions were to give the AI problems it would make mistakes on. I was surprised at how quickly and easily the AI solved many word problems I'd rate as medium or difficult. I did eventually get it to make an error in its reasoning by presenting it with the Collatz Conjecture, though the AI recognized the conjecture in its final reply.
I also managed to get to make an error with a question involving the coin rotation paradox. A similar question was on the 1982 SAT, and everyone, including the test makers, got it wrong because the correct answer was not among the choices. Later, I notice the model struggles with questions involving complex numbers. That's not surprising since such numbers are a somewhat obscure topic with limited applications. It also struggled with binomial expansions and large exponents.
Overall, I was very impressed with its reasoning ability. It can solve math problems that are posed without explicit equations. So in addition to understanding math, it also has good English comprehension.
There is an ongoing debate about the limits of AI. At this point, I'm convinced it will continue to advance, which is an exciting possibility, though many people will be forced to find new jobs. Even so, there will always be jobs for humans, either because machines can't do them, or people are better at them.
My advice is to relax and enjoy the show. In the end, there are no rules, and the points don't matter.
No comments:
Post a Comment