Reinforcement Learning and AI Agents
In Chapters 5–7, we discussed how to provide our model with access to external memory. This memory was stored in a database of a different type (vector or graph), and through a search, we could look up the information needed to answer a question. The model would then receive all the information that was needed in context and then answer, providing definite and discrete real-world information.
However, as we saw later in Chapter 7, LLMs have limited knowledge and understanding of the real world (both when it comes to commonsense reasoning and when it comes to spatial relations).
Humans learn to move in space and interact with the environment through exploration. In a process that is trial and error, we humans learn that we cannot touch fire or how to find our way home. Likewise, we learn how to relate to other human beings through interactions with them. Our interactions with the real world allow us to learn but also to modify our surroundings...
