You can get better code by exploiting model weights (kelvinfichter.com)

A recent exploration into AI model behavior reveals significant insights on how naming conventions impact a model's understanding of systems, particularly in software development. The author recounts an amusing episode with the AI Claude, which, when asked to estimate the duration of a complex software refactor, predicted seven weeks—far longer than the actual six hours required. This discrepancy highlights that AI models, trained on vast datasets comprising human project experiences, can misapply learned priors, such as equating lengthy refactors with complexity, rather than recognizing their own efficiency. This observation underscores a broader implication for the AI/ML community: by aligning software components with well-understood domains—like legal or military systems—developers can enhance how models intuitively grasp the architecture and dynamics of their applications. Metaphorical naming provides models with an architectural advantage, enabling them to leverage existing knowledge rather than starting from scratch. While this strategy holds considerable promise, users must remain vigilant about the limitations of metaphors, as mismatches can lead to confusion. As AI continues to evolve, this technique may offer a temporary edge in harnessing models' strengths, allowing developers to achieve better outcomes with less explanatory effort.