Google DeepMind's Haydn Belfield argues that "tokenmaxxing" leaderboards are primarily useful for testing AI model limits · Digg