AI Coding will Prevent Expertise

gist

Lars Faye argues that AI coding tools demand the expertise they also prevent novices from forming. Veterans who already have taste can steer and audit models. Newcomers are told to use assistants that skip the friction that builds that judgment. Studies of novice programmers and students show heavy assistance creates an illusion of competence, while people who treat models as Socratic tutors keep the cognitive work. Faye's prescription is to treat generation as a leaky abstraction and keep friction first: documentation, exercises, and verification against sources other than the model.

ideas

• The expert-novice bind is structural. Coding assistants require taste and review to use well, and they also remove the friction that produces those skills.
• Unrestricted generation produces an illusion of competence. In the JetBrains-cited novice study, heavy Copilot users skipped planning and finished confident without understanding. People who ignored bad suggestions wrote the code they already intended.
• Learning with LLMs is inverted. The student has to steer the mentor first, so domain knowledge is what makes the answers checkable. Without it, the model confirms whatever direction the prompt already implies.
• Friction is the training. Taste and fingertip feeling come from failing, tracing, and rewriting. Socratic use keeps that work on the person; answer generation outsources it.
• Treat generation as a leaky abstraction. Use models for interactive docs and exercises, distinguish cognitive debt from offloading, and verify claims against documentation, peers, and trial and error.

quotes