Your analogy to directed evolution in laboratory settings is apt. We understand enough to guide the system toward desired traits, but the intricate details of these changes are not fully grasped. Neural networks, including large language models (LLMs) like GPT-4, operate on fixed capabilities once their training is complete. This 'checkpoint' represents a snapshot of the network's abilities, which remain static until manually updated. Alterations to the network, such as layer adjustments or addi...