The author criticizes the GPT 6.1 model for being slow, unreliable, and poor at instruction-following, suggesting a return to earlier GPT versions like 5.6 or 5.4. They propose external accountability controls to manage changes and violations in usage.