AI Can Help You Code Faster – But at What Cost (codesmarternotharder.substack.com)

A developer recounts how generative AI tools—starting with ChatGPT and Claude and now agentic IDEs like Cursor (Cursor Tab can watch keystrokes and refactor or generate multiple files from a single prompt)—boosted their productivity by an order of magnitude. Using prompts, detailed context, and iterative validation, they shipped a Golang cryptocurrency-payments gateway integrating three exchange APIs and now rely on tools like CodeRabbit for local commit reviews. The workflow moved from digging in docs and Stack Overflow to instant, often correct suggestions that sped up shipping and scaled throughput. But that speed came with a cost: skill atrophy and erosion of critical review. Banned from AI during a LeetCode-style challenge, the author couldn’t recall basic TypeScript loop syntax or JavaScript array helpers and failed an order-book calculation they’d once solved easily. The piece warns that agentic systems’ confident answers can lull engineers into rubber-stamping code, weakening collective critique and the ability to handle novel, long-horizon design problems. For the AI/ML community the implications are concrete: design tools and benchmarks that preserve human oversight (uncertainty estimates, explainability, human-in-the-loop workflows), encourage periodic manual practice, and treat generative systems as amplifiers—not replacements—of human engineering judgment.