Coding With AI Is Like Driving A Race Car

Giving a powerful AI coding agent to someone who can’t read the code is like putting an untrained driver in a race car. The problem is not the car. The problem is that nobody at the wheel knows the track.

The car is genuinely fast. AI writes hundreds of lines of working code in the time it used to take to write ten. That is real, and it is not the thing that gets you hurt. Speed with no track knowledge does not win races. It puts you into the wall on the first corner you didn’t see coming.

I learned what safe crashing is worth at 15

When I was 15 I joined my school’s robotics team. I had six months to program a robot to follow a black line on the floor. I failed. Completely. Not “it worked in practice but not on the day” failure. It never really worked. My teacher watched me fail for six months and never once stepped in to fix it for me. He let me take that broken robot all the way to the state competition, where it stalled and did nothing in front of everyone. I felt ashamed standing there.

Him not saving me was the entire point. That was the cheapest crash of my life. No client, no money, no one depending on the robot doing anything. Just a teenager learning that a plan that looks fine in your head can stall in front of a room full of people, and that this is survivable. Every developer I respect has a stack of crashes like that behind them. That stack is the repertoire. It is the only thing that tells you, before you commit, which decisions look clever now and wreck you in three months.

The failure mode is not “AI writes bad code”

Everyone worries about the wrong risk. The fear is that the AI will produce garbage. It usually doesn’t. The harness holds, the code runs, the demo works. The real failure mode is quieter: a person with no repertoire lets the AI make the decisions.

AI is a tool. We are just not used to tools that talk back. A hammer never proposed an architecture. A compiler never suggested that you store passwords in plain text because it was faster. The agent will, in a confident and well-formatted paragraph, and if you can’t read what it produced you have no way to push back. So you accept it. Not because you agreed with the decision. Because you didn’t know there was a decision being made. That is the driver taking their hands off the wheel because the car seems to know where it’s going.

The cost is not the same on a side project and on a paying client

A non-technical person can absolutely learn to drive this. I am not telling anyone to stay out of the car. I am saying learn where crashing is cheap.

The right place to crash is a side project. No paying clients, no expectations, nothing riding on it but your own time. That is exactly how today’s developers built their judgment. We shipped broken things into the void for years before anyone trusted us with their money. The barrier to entry did not disappear because the AI showed up. It moved to a place people can’t see, which is the repertoire you only earn by getting things wrong somewhere safe.

I know what the other version costs because I lived it. I started a startup with a former colleague. We had a vision we believed in, real money going out, and real time we were never getting back. When early signals told us people didn’t want the thing, we decided the idea was fine and our pitch just wasn’t clear enough. We drove that conviction straight into the wall. That crash was not cheap. It taught me the difference between a decision you get to redo and a decision that costs you. Taking your hands off the wheel on a product real people paid for is the expensive kind.

The barrier to producing software is gone. The barrier to producing software that survives contact with real users is exactly where it always was: judgment, earned by crashing where it was safe to crash. Get in the car. Learn the track on the days nobody is racing.