self prompting via chain of thought and tree of thought can be used in combination with updating memory containing knowledge graphs combined with cognitive architectures like SOAR and continuous external new information and sensory data … with LLM at the heart of that system and it will exactly be a “thinking machine”. The problem is currently it’s very expensive to be continuously running inference full time and all the engineering around memory storage, like RAG patterns, and the cognitive architecture design is all a work in progress. It’s coming soon though.