The Dangerous Illusion of AI Coding? - Jeremy Howard

Howard’s core claim: AI coding tools create the illusion of velocity. Teams feel 10–50x faster, but shipped, maintainable value barely moves. His group’s study reports only a tiny uptick in actual output. The gap is explained by “vibe coding”: stochastic trial-and-error that looks productive, rarely survives integration, and quietly increases tech debt.

He distinguishes coding from software engineering. LLMs interpolate across known patterns and pass tests when the path is well-trodden, but they reliably collapse outside the training distribution. Without strong constraints (tests, verifiers, specs), you get opaque code nobody understands. Agentic flows amplify the slot-machine dynamic: you tweak prompts, pull the lever, and occasionally “win,” reinforcing a process that undermines systematic design and diagnosis.

Organizationally, the danger isn’t unemployment—it’s skill atrophy. Seniors benefit (faster typing, search, boilerplate); absolute beginners can ship toy apps. Mid-level growth stalls: less hands-on design/debug experience, weaker mental models, and mounting “understanding debt.” Over time, this erodes the firm’s capacity to reason about systems—exactly the muscle you need for novel work and reliability.

The workable use-cases are narrow: small, fully understood changes; code you can specify crisply; areas with dense, trustworthy test harnesses and fast feedback loops. Treat LLMs as assistive autocomplete and research copilots—not architects or owners of ambiguous features. Invest in tests over prompts, and enforce human design reviews, code ownership, and traceability.

Hype to discount: claims of clean-room originality from models, sweeping productivity multipliers, and near-term agentic autonomy. Evidence remains anecdotal; even Howard’s study is referenced, not detailed. Leaders should demand outcome metrics tied to business value, incident rates, and maintainability before scaling AI-first development processes.