Coding Is Not Solved – Alex Ewerlöf Notes

An experienced engineer argues that despite flashy demos, coding is far from solved by LLMs. The core claim is that generation is cheap but the hard parts - non-functional requirements like maintenance, reliability, security, scalability and accountability - remain human problems. LLMs are stochastic and brittle with logic and large contexts; they succeed only when wrapped in traditional engineering practices (harnesses, tests, chain-of-thought, agents) that hide or repeatedly correct errors. Some uses tolerate unread code (personal projects, proofs-of-concept, or deliberate weaponization), but sectors where mistakes cost money, lives, or legal exposure - healthcare, finance, automotive, aviation, power, defense - require human accountability that AI cannot provide.

The piece warns against managerial pressure to bake AI into every workflow and against complacent narratives that conflate motion with progress, English specs with code, or aesthetic “taste” with engineering value. It catalogs common fallacies, outlines pragmatic mitigations already in use, and laments service degradation when vendors prioritize velocity over verification. It calls out “AI overdose” - skill atrophy and poor judgment - and highlights dangerous shortcuts like massive PRs, parallel agents, and looped prompts that obscure defects. The practical takeaway: use AI as a force-multiplier for specific tasks, but insist on rigorous verification, measurable service levels, and continuous upskilling to remain accountable and relevant.