Quoting Bryan Cantrill

Cantrill's framing shifts the conversation from capability to incentive structure: the problem isn't that LLMs write bad code, it's that nothing in their design penalizes waste the way a human engineer's finite hours do. That's a different critique than most quality-focused objections, and it has organizational implications that pure benchmark comparisons miss.

This connects directly to TechCrunch's 'Tokenmaxxing' piece from earlier this week, which documented a concrete instance of exactly the dynamic Cantrill describes: developers optimizing for apparent productivity while accumulating maintenance debt that offsets the gains. Together, the two pieces suggest a pattern where the incentive misalignment isn't incidental but structural. The MIT Technology Review argument about treating enterprise AI as an operating layer is also relevant here — if the competitive advantage lies in governance and refinement infrastructure rather than raw model output, then bloat-prone codebases become a liability at the infrastructure level, not just the sprint level.

Watch whether engineering organizations that have publicly committed to AI-assisted development — Shopify and Duolingo are recent examples — start reporting measurable increases in codebase complexity or refactoring costs within the next two to three quarters. That would be the first empirical confirmation of Cantrill's structural claim at scale.