MIT researchers have found a mechanistic explanation for why larger language models reliably get better.

Source Brief

MIT researchers have found a mechanistic explanation for why larger language models reliably get better. The phenomenon, called 'superposition', shows how neurons store multiple concepts simultaneously.