Simon Willison's post surfaces a critical, second-order question for AI tooling: where does automation actually create more work? Varda's reported experience suggests that for tasks like writing commit messages, which require synthesizing intent and impact, current AI may create a net-negative feedback loop. The output demands human correction, adding a step rather than removing one.
This indicates that the industry's push for automation may need to more carefully map which developer workflows are truly ripe for augmentation versus those where AI currently introduces noise. The real test will be whether future models can learn to prioritize the 'why' over the 'what'.