The model providers are still training the models to know how to do arithmetic.
Only if you're Seeing Like A Frontier Lab would you do it that way.
That's insanely wasteful.
We can write mechanistic code that does that calculation right every time!
LLMs can distill that mechanistic code themselves!
Only if you're Seeing Like a Frontier Lab would you try to reduce absolutely everything to a single gradient, no matter how grotesquely expensive the compute it requires.
The more that mechanistic code can do the calculation, the less power the Frontier Labs have over you.
It makes sense they want everything to be done by a single god model that is insanely expensive to create and thus has few competitors and they can charge whatever they want.
Everyone else should want intelligence to be commodity, not proprietary.