Rules don't scale. Data does.

For years, the way to make a car drive itself was to write rules. Stop at the stop sign. Stay under the speed limit. Yield to the car on the right.

Tesla's self-driving software had more than 300,000 lines of this kind of hand-written code.

Then Tesla threw nearly all of it away.

The problem with rules

Rules aren't scalable. Every rule has exceptions, and the exceptions have exceptions that nobody can program in advance.

A traffic cone in the lane. An officer waving you through a red light. A pothole the size of a kiddie pool. When the car meets a situation nobody wrote a rule for, it behaves in unexpected ways.

Drive like we drive

With version 12 of its Full Self-Driving software, Tesla replaced those rules with a single neural network trained on millions of video clips of people driving. It also bet on cameras alone, dropping the other sensors.

The idea is to drive like we drive. People don't consult a rulebook at a four-way stop. We've watched it done ten thousand times.

That approach needs massive computing power and very large data sets. Tesla has the second because it has millions of cars on the road sending back video.

The result is a car that slows down in bad weather, steers around cones, pulls over for emergency vehicles, and avoids potholes. All of those are extremely difficult, or impossible, to program one rule at a time.

Three caveats

• It still needs a driver. Tesla's own name for the product is "Full Self-Driving (Supervised)."
• Learned systems are harder to explain. With rules, you can find the line that caused a mistake. With a neural network, you often can't say why it did what it did. That makes testing and oversight more important.
• The race isn't settled. Competitors pair learned systems with extra sensors and detailed maps, and they have driverless cars on the road too.

Why this matters if you don't make cars

Every organization has rulebooks that hit the same wall. Eligibility rules. Routing rules for service tickets. Fraud flags. Document sorting. Each one gathers exceptions until nobody fully understands it.

Learning from examples scales where rules can't, provided you have the data to learn from.

That's the larger point. A company's competitive advantage will increasingly be tied to the AI models it builds and maintains, and to the data behind them.

Tesla's bet was to stop writing better rules and start collecting better examples. The question for the rest of us is which of our own rulebooks should be next.

A shorter version of this article first appeared on LinkedIn. Join the discussion there.