The truth about productivity gains in software with AI

AI tech bros announce a 2x to 3x productivity gain thanks to AI in the field of software engineering. A few days ago, I read a study on the subject that showed a 10% gain.

Why this difference?

Lies and aggressive marketing?

Not only.

On my side, when I start a new project, without having to interact with legacy code, my ability to produce code is effectively multiplied by 2 or 3.

Here I am not talking about producing code massively by relying on agents. Otherwise, my productivity would be multiplied by 20. But the result would be of little value.

I’m talking about creating quality code. I asked the AI to create the code, but then I take the time to proofread it, make it my own, organize it, and improve it.

Basically, AI is like an intern that I have to guide step by step. And I take on the roles of lead dev, architect and staff engineer at the same time.

Of course, the comparison between AI and an intern is very bad. We must not forget that AI has nothing in common with a human.

Even if it gives the illusion of it, AI doesn’t understand anything. It is just able to provide a probable answer to our questions, based on statistics. And if its training data covers the subject, it will be able to answer without hallucination.

It’s a bit troubling because unlike an intern, the AI can know all the doc and all best practices. But this does not mean that it would be able to apply them in a concrete case.

That’s why we must never forget that AI doesn’t work like a human.

Let’s get back to the original subject. Why this difference?

Well, as soon as I have to interact with legacy code, my productivity drops and the advantage of using AI becomes quite small.

To say that AI can multiply our ability to produce code by 2 or 3 is true. But to stop at this statement is to forget that the main job of a software engineer is not to create code. The value is not here.

Most of the time, we do maintenance. We read code to remember it. We think. We search. We organize. We debug.

And for all these tasks, even if sometimes, AI can help us, it only saves us a little time.

Would it be possible to change our working methods to gain much more than 10% productivity?

I think so.

In the corporate world with projects that are ten years old or more, it’s complicated to become AI friendly overnight.

But what we need to understand today is that producing code has become cheap. On the other hand, engineering is still as expensive.

It’s really easy and fast to produce a unit of 2000 lines of code with AI. More than ever, it is necessary to organize our code into standalone modules. While this principle is nothing new, it has never been more important.

When we say this today, people tend to think of services. But depending on the case, it can also be libraries, or object programming.

And when I talk about object programming, I think of Alan Kay’s vision, not the delirium of abstraction that we learn at school, because even if it can be intellectually stimulating, the result is difficult to maintain and debug. Academic OOP leads to a strong coupling and this is the opposite of what we want.

Organizing code into small, independent units allows you to make maximum use of the AI’s leverage. Because each piece of code is like a new project, which can be created quickly, or rewritten from scratch as it evolves.

And I am convinced that with a project designed from the beginning for AI, it is possible to increase productivity by much more than 10%, even after years of history.

Don’t miss my upcoming posts — hit the follow button on my LinkedIn profile