Models Spend Same Compute on Every Token

AnneliesGamble posted that current models assign the same compute to each token regardless of whether the token is easy or hard. The post references a conversation with Prateek Jain, Distinguished Scientist at Google DeepMind India who co-leads long-term model research for Gemini. Jain retweeted the post. The visible text stops at the start of what the two discussed and supplies no further details on any proposed changes or outcomes.