Axiom's new research blog: Learning Collatz - The Mother of all Rabbit Holes What happens when you train a transformer on one of math's most infamous unsolved puzzles -- and then study how it fails?