Jeff Smith's QCon London 2026 talk addresses the growing mismatch between AI coding models and real-world codebases. AI tools are trained on public repository snapshots and lack organization-specific knowledge, causing them to generate syntactically correct but architecturally non-compliant code. Data shows AI-assisted pull request volume has surged since 2022, but acceptance rates have dropped. Smith categorizes repository constraints into architectural rules (component patterns, dependency handling) and procedural rules (PR conventions, testing expectations), noting these are often enforced implicitly during code review. He proposes 'repository fingerprinting' to systematically extract and document these implicit rules, making them accessible to AI systems. The core argument is that the AI-code quality gap is a knowledge management problem, not a tooling problem, and teams that explicitly document their repository constraints will gain a competitive advantage.
Sort: