Appendices

Agents

Large language models are not artificial intelligence in the classical sense of a system that holds a model of the world and adjusts it; they learn, but not dynamically, being pretrained. The recent innovation is the agent: a model given sensors and actuators by a harness, which is a sampling apparatus in exactly the sense the chapter on knowledge gives the word. The claim worth making is that the intelligence is the fine-tuned agent rather than the base model – base models encode knowledge and regurgitate it, and intelligence emerges when a model is fine-tuned to act within an environment, which is also why “thinking models” are better described as having learned how to act than as merely thinking more. The note this is drawn from is parked for later writing.