DeepMind is reportedly developing a system that gives AI agents an “inner monologue” to help them learn tasks more efficiently. In a recent patent filing, Google’s AI lab outlined a method called “intra-agent speech to facilitate task learning,” where robots observe tasks through images or videos and generate natural language descriptions of what they see. This internal narration helps link visual input to actions, enabling “zero-shot” learning and allowing agents to understand and interact with unfamiliar objects without prior training while also reducing memory and computing demands. Last week, in a breakthrough for genomics, DeepMind introduced a powerful AI model capable of predicting how individual DNA mutations impact the intricate systems that control gene activity. DeepMind is exploring a new way to help robots learn by giving them an “inner monologue”. According to The Daily Upside, in a recent patent filing, the Google-owned AI lab detailed a system that allows AI agents to observe videos or images of a person completing a task and then generate natural language descriptions of what they see. For instance, a robot might watch a video of someone picking up a cup while internally processing the phrase “the person picks up the cup.”

Full report : Google DeepMind’s new AI lets robots learn by talking to themselves.