Coding assistants are solving the wrong problem
a day ago
- #AI coding assistants
- #technical debt
- #software development
- AI coding assistants increase task completion by 21% but show no improvement in company-wide delivery metrics.
- Experienced developers are 19% slower with AI assistants but believe they are faster.
- 48% of AI-generated code contains security vulnerabilities.
- Developers' main job is to reduce ambiguity, but AI assistants often increase it.
- AI assistants require clearly-defined requirements but often bury requirement gaps in code.
- AI-generated code leads to more downstream code reviews and security patches.
- Seasoned developers benefit from AI by focusing on architecture while AI handles implementation.
- Junior and mid-level engineers face increased pressure to ship faster with unreliable AI output.
- Only 16% of a developer's time is spent writing code; the rest is operational work.
- AI assistants save 10 hours per week but inefficiencies elsewhere cancel out gains.
- Technical debt is often created in product meetings, not in code.
- Developers frequently discover unexpected codebase constraints after committing to a product direction.
- Key desired improvements: reducing ambiguity upstream and clearer picture of affected services and edge cases.
- Most costly defects stem from misalignment between requirements and architecture.
- LLMs can better map existing code structures than generate fully-functional code.
- Developers are open to tools that augment workflows but want flexibility in deployment.
- Bicameral aims to deploy AI pragmatically, focusing on human needs and context-sharing.