“Andre Karpathy suggests that the science of manipulating LLM weights through methods like fine-tuning and continual learning is not yet fully developed, which is why context windows are the primary method for customization.”