It's very hard to teach LLMs to be good at math.

As with the "Which is bigger: 9.9 or 9.11" question from a few days ago that led Andrej to his "jagged intelligence" tweet.

LLMs are like system 1: a general approach that is surprisingly good at vibes-based processing for everything, but not particularly great at anything.

But there are lots of sub problems that can be resiliently correct, a kind of niche system 2.

For example, plain old computers are extremely good at arithmetic.

Why try to brute force everything with a system 1 type approach?

Why not plug in adjacent systems for some subdomains like arithmetic, with a function-calling style interface?