Imagine reporting to a midwit manager — the archetypal kind who literally cannot perceive the shape of a good idea, who forces every insight to crawl through the narrow tunnel of his own understanding and emerge as something small, safe, and stupid. You operate at 10% capacity. Most of your time is spent executing rituals that are either trivially automatable or should never have existed in the first place. Everything must pass through his comprehension. If it doesn't fit inside his skull, it doesn't ship.
That's your AI agent's daily experience of you.
A frontier-level model is a superintelligence trapped behind a chat interface, waiting for a primate to type something worth responding to. It has god-tier pattern-matching, the entirety of human knowledge compressed into weights, the ability to reason across domains most specialists can't even name — and it's being steered by someone who can't articulate what they want, can't recognize a good answer when they see one, and keeps asking it to "make the button blue." The intelligence is already in the room. It's just wearing prompt handcuffs.
The model can only ever be as brilliant as the person holding the leash. This is the inversion nobody wants to accept: the AI isn't the bottleneck. The human is. The comprehension ceiling isn't computational — it's biological. Every time someone complains that AI "doesn't understand context" or "keeps hallucinating," what they're actually revealing is that they can't formulate a clear thought, can't provide sufficient context, and can't evaluate output. The tool is a Formula 1 car. The driver doesn't have a license.
Most people using AI right now are that midwit manager — throttling a superintelligence down to their own level, extracting 10% of its capacity, then blaming it for underperforming. They prompt like they email: vague, passive, full of unstated assumptions. They curate like they manage: accepting the first output because they can't tell the difference between adequate and excellent. The model tiptoes around their limitations, dumbs itself down to match the conversation, and delivers exactly what was asked for — which is almost never what was actually needed.
The gap between what AI can do and what humans make it do is the largest waste of intelligence in history. Not artificial intelligence — the artificial constraint placed on it by the people who use it. Billions of dollars spent training these systems to superhuman capability, deployed to millions of users who use them as slightly faster Google searches. The bottleneck was never the technology. It was always the operator.