An AI Coding Analogy

2 months ago

In watching the AI hype cycle/bubble play out, it's become clear that using AI (LLMs) for writing code is a difficult topic for people to reason about because the behavior and nature of LLMs is so foreign and new to most people. So, I came up with the following analogy for AI coding in an attempt to help clear up confusion for folks.

Traditional coding is like driving a car – AI coding is like being a passenger in a car and telling the driver what to do.

The driver doesn't actually know how to drive, but they have watched every video of every driver ever and can mimic them. The driver may sometimes safely reach the destination without assistance. Other times, the driver gets lost or attempts to drive off a cliff. You never know what the driver will do next but they do drive really fast.

The passenger may have a lot of driving experience, some experience or none at all. The passenger is not driving, they are not gaining driving expertise, and they are at risk of losing what driving expertise they may have.

They may reach the intended destination in one piece - this time. But the more complicated, longer and less well-traveled the route to your destination is and the less driving expertise the passenger starts off with, the more likely they fail to arrive at the desired destination.

The passenger could micromanage the driver to the point where the success rate of any trip is pretty high - but it likely would have been easier if they just drove themselves.

The passenger could fill the car with more of these types of drivers and have them all work together to drive. But it doesn’t materially improve the odds of reaching the destination because every one of them makes mistakes the same way.