The original title is "When AI Beats a Senior Developer — and When It Doesn't"

Short summary

A senior developer's framework for deciding which tasks to delegate to AI agents: the key distinction is interpolation vs extrapolation. AI excels at tasks where the answer lives densely in training data — boilerplate, format translation, well-trodden patterns. It fails on tasks requiring extrapolation beyond training data, verification against unseen reality, or consequence-modeling — like a multi-tenant billing query where the invariant exists only in the engineer's schema and head.

• •AI wins on interpolation tasks (boilerplate, format translation, common patterns) and loses on extrapolation, verification, and consequence-modeling
• •Real example: agent wrote a 200-line normalizer in 90 seconds but produced a subtly wrong multi-tenant billing query
• •Framework: ask whether the task is inside well-covered training space or requires reasoning beyond it

Generated with AI, which can make mistakes.