The expert approaches every problem thinking “how would I build this?” — gets stuck on implementation details, spends hours on things AI handles automatically. The novice describes the outcome they want, lets the agent figure out implementation, and ships in minutes. (Note: this expert-vs-novice framing synthesizes themes from multiple sources rather than quoting any single one directly.)
This is the counterintuitive implication of the implementation gap collapsing. It doesn’t mean technical knowledge is useless — it means the reflex to think in implementation terms actively slows you down. The shift Karpathy describes in Declarative beats imperative when working with agents is the same insight from the engineering side: stop telling, start specifying.
Karpathy himself noted his manual coding ability is already atrophying, and predicted a split in engineers: builders will thrive, pure coders may struggle. Boris Cherny confirms the endpoint: Every role codes when implementation cost drops to zero — the generalist builder replaces the specialist engineer — at Anthropic, PMs, designers, and the finance team all code because AI handles implementation, making the “software engineer” title vestigial.