Three Limits Still Apply
AI tools can accelerate useful parts of software work: search, summarization, drafting, boilerplate, test scaffolding, refactoring suggestions, and first-pass code generation. They do not remove the basic limits that make software engineering hard.
No Free Lunch still applies: a method that works well for one class of problems can fail on another. No Silver Bullet still applies: requirements, state, change, integration, and domain structure do not disappear. The Mythical Man-Month still applies: adding more apparent labor can increase coordination and review cost instead of reducing it.
Where AI Helps
AI is useful when the task has enough context, enough examples, and a verification path. It can speed up repetitive edits, summarize logs, propose tests, draft documentation, and help search a large codebase.
The useful standard is not whether output looks plausible. The useful standard is whether the result can be checked against the repository, runtime behavior, tests, product requirements, user data, or the physical artifact being described.
What Still Requires Engineering
- Requirements: deciding what the system must do for real users and operators.
- Architecture: choosing boundaries, state ownership, data shape, deployment shape, and failure behavior.
- Integration: fitting changes into existing code, migrations, permissions, workflows, and support paths.
- Verification: proving behavior with tests, traces, rendered output, measurements, or production backports.
- Maintenance: keeping the result understandable after the first generated answer is gone.
Useful Scope
This note is not anti-AI. It is a boundary note. AI assistance is strongest when the work remains attached to source material, checks, constraints, and responsibility for the finished result.