21 hours ago
- AI agents cannot truly program; they are statistical models that produce increasingly hard-to-detect broken code.
- The author experimented with agents for 6 months but found manual work more efficient; agents frontload progress but fail at polish.
- AI is useful for prototypes and better than Google for searches but is not a substitute for software engineers.
- High-performing individuals error-correct well with agents, but low-performing ones in large organizations produce more 'slop' without self-checks.
- The assumption that AI-generated artifacts are created by a human-like process leads to fragile, unbuildable code.
- The author aligns with LeCun/Marcus view that true programming agents need world models, not current RLVR approaches.
- The main risk of AI adoption in software is large organizations harming themselves in their 'AI psychosis'.