They experimented with gpt-2 scale models. Hard to make any meaningful conclusions in the gpt-5 era.