You've probably seen the videos going around where someone asks an AI whether...

You've probably seen the videos going around where someone asks an AI whether they should walk or drive to the car wash. The car wash is 50 yards away. The AI tells them to walk.

To the car wash. (With their car.)

This is Opus 4.7. The same model that can refactor a 100,000-line codebase and find security holes nobody else caught. State of the art. Can't figure out that you need your car at the car wash.

Andrej Karpathy (the guy who coined "vibe coding") shared this example in a recent keynote. He calls these models "jagged." Spectacularly good inside the circuits the labs trained them on. Unreliable outside those circuits. And nobody gives you a map of which circuits you're in for your specific use case.

He shared a quote he says he comes back to constantly: "You can outsource your thinking but you can't outsource your understanding."

That lands differently when you're running a business. Automation multiplies whatever direction you point it. If nobody on the team actually understands the problem you're solving, the agent will confidently build the wrong thing at 10x speed.

Karpathy is still the bottleneck in his own projects. Not for writing code (the agents handle that), but for catching mistakes only a human would notice. His coding agent tried to match users by email address across two services instead of using a persistent user ID. Obvious if you understand the system. Invisible to the model.

Everyone can build now. But the ceiling still belongs to people who understand the problem well enough to know when the output is wrong — even when it looks right.

That's the skill that compounds. Not prompting. Understanding.