Pure-mathematics background, working on AI safety.

Selected work

• How a network learns a finite group: irreps vs cosets, tested on a same-character-table pair

  Does a grokked network compose group elements through their irreducible representations or through cosets? Two groups with identical character tables, built to tell the accounts apart, trace what divides them to representation type: the coset account adds nothing the irreps don't, while the real-versus-quaternionic split shows up twice over — in how fast each group learns and in a matrix-level readout.

More about me and what I'm looking for →