Hi, cool project! Which language models are you using for the summaries?

Quite a few actually. We will do a writeup once everyting is stable and post it on HN

I've tried this

> 19999999999999 + 100000001 > The sum of 19999999999999 and 100000001 is 20000000000000.

Which is incorrect. I guess it just returns the most likely result.

Most likely result according to what? If its never seen those numbers before how would it know?