According to CoinMeta, Anthropic published a study regarding its AI model Claude, which introduced the “dreaming” ( dreaming ) mechanism of Claude. This mechanism allows Claude to review historical records during task breaks, correct mistakes, and extract lessons. The study utilized approximately 350,000 internal company documents, totaling 420 million token, and after 128 rounds of conversations, the问答 scores of Claude increased from 53% to 61%. In addition, Claude also participated in the game “Pokémon Red”; the new version of the model is better at distinguishing strategies that are worth saving and game states that can be re-examined at any time.