Why 'Coding Is Solved' Is Wrong
Original: Coding Is Not Solved
Why This Matters
The AI coding hype cycle is colliding with production reality; this debate shapes hiring, tooling, and liability decisions across engineering.
Engineer Alex Ewerlöf argues that LLMs have not solved coding, citing accountability gaps, non-functional requirements, and logic constraints that current AI tools cannot reliably handle. His post targets the narrative spreading in tech circles that software engineering is now just a matter of 'taste.'
Alex Ewerlöf, a veteran developer with degrees in hardware and systems engineering, pushes back hard on the popular claim that LLMs have made coding a solved problem. His core argument: code creation is cheap now, but that was never the expensive part. Maintenance, reliability, security, and scalability — non-functional requirements (NFRs) — represent the bulk of real engineering costs, and AI handles those poorly.
He carves out three narrow cases where not reading AI output is acceptable: personal automation, proofs of concept, and what he bluntly calls 'weaponized AI.' Everything else — healthcare, finance, aviation, defense — demands accountability that AI structurally cannot provide. You can't fine a model. You can't imprison it. Without consequence, there's no accountability.
On the technical side, Ewerlöf points out that coding is fundamentally about logic, not language. The reason AI agents appear to work is because engineers have built feedback loops that feed compiler and runtime errors back into the model repeatedly. That's error-masking, not competence. He also notes Anthropic's own Claude Code leaked with multiple flaws, and that Anthropic's status page has been showing persistent service issues — orange is the new green, he quips.
His sharpest shot: the loudest voices claiming 'coding is solved' tend to have nothing running in production at scale to show for it.