Charity Majors argues that AI development requires stricter engineering discipline rather than a looser approach, countering the narrative that rapid iteration should replace traditional software practices.
Majors, an infrastructure and observability expert, frames the debate as asymmetrical. AI enthusiasts rush to deploy systems before understanding failure modes. AI skeptics worry about systems degrading over time without proper maintenance and oversight. The real problem lies not with skepticism but with abandoning proven engineering practices in pursuit of speed.
The core tension: machine learning introduces new failure modes that traditional software engineering doesn't fully address. A typical application either works or fails in predictable ways. An AI system can appear to work while producing subtly incorrect outputs that compound across deployments. Users may trust results they shouldn't.
Majors emphasizes that mature engineering disciplines exist for good reasons. Code review, testing, monitoring, and gradual rollouts catch problems before they scale. These practices slow development initially but prevent catastrophic failures later. AI teams often skip these steps, assuming continuous learning will fix errors automatically. It won't.
The article pushes back against false choices. Building AI systems responsibly doesn't mean avoiding deployment or rejecting innovation. It means applying observability, testing frameworks, and staged rollout strategies that software engineering developed over decades. These practices protect users and maintain system reliability.
Majors's underlying argument resonates with infrastructure professionals who've watched systems fail when shortcuts replaced discipline. AI isn't magic that transcends engineering fundamentals. It requires the same rigor as any critical system, plus new tools to handle model-specific risks like drift and hallucination.
The stakes matter here. As AI systems make decisions affecting hiring, credit, healthcare, and security, cutting corners on engineering discipline creates real harm. Speed without structure produces brittle systems. Building trust in AI means committing to the unglamorous work of monitoring, testing, and incremental improvement