‘AI’ Sucks the Joy Out of Programming (alexn.org)

A long‑time programmer reports that LLMs and agent-driven tooling have drained the joy from programming by replacing the learning journey with brittle, frustrating automation. In practice the author finds models and IDE/CI agents handle the “boring” surface work—API wiring, scaffolding—but repeatedly fail on hard, algorithmic or concurrency problems. Prompting an agent to iterate is slow and error‑prone: it often fixes superficial issues, compounds subtle bugs, and produces code that initially looks fine in PRs but grows into unmaintainable “crap” with poor comments. The result is stress, loss of control when debugging non‑deterministic performance/concurrency issues, and a diminished sense of mastery when things finally work. For the AI/ML community this is a cautionary note about current limitations of code‑generation systems and agent workflows. Key technical implications include model unreliability on complex reasoning tasks, weak debugging and iterative correction, accumulation of technical debt in generated code, and a fragile human–model feedback loop. The piece underscores needs for better grounding, verifiable outputs, stronger developer tooling (automated tests, provenance, explainability), and design choices that preserve learning and maintainability rather than just automating surface tasks.