Why Your AI Never Works on the First Try

The article argues that repeated AI revisions are not a bug to be fixed with better prompts, but a constraint of how LLMs work. They move one token at a time from the current position, so they can get closer to a target only through iteration, error messages, and human guidance.

It frames AI as a cognitive exoskeleton rather than a replacement for human intelligence. The user provides the map, judgment, and quality control, while the system provides raw generation and pattern matching. That makes iteration unavoidable, but also useful when the task justifies the effort.

Reading notes

#

• AI work often requires many attempts, even at temperature 0, because the model cannot see the full destination from the start.
• The pattern appears across code, email, and images, especially when the user knows what good output should look like.
• Iteration works like stepping through semantic space one intersection at a time, often from error messages or corrections.
• Better prompting, multiple agents, and elaborate workflows do not remove the underlying limit.
• The article compares LLMs to a traveling salesman without a map and links the problem to online graph exploration and partially observable planning.
• Expertise helps because it gives the user a mental map of quality and lets them steer revisions more effectively.
• The article calls AI a cognitive exoskeleton: it amplifies human cognition but depends on constant human control.
• Some tasks are worth grinding through; others can stay at “good enough.”