Anthropic published a very interesting article on recursive self-improvement: the possibility that AI systems will increasingly contribute to building better future versions of themselves.
Article: https://www.anthropic.com/institute/recursive-self-improvement It is a huge topic, almost science fiction: AI developing AI, increasingly fast improvement cycles, systems capable of automating growing parts of research and development.
But reading the article as a developer, I think the most interesting question is not:
Will AI replace us?
The more useful question today is another one:
If AI can write, test, refactor, and review code better and better, what is the real role of humans in software development?
For years, we have measured a developer's work partly by their ability to produce code: building features, fixing bugs, optimizing parts of a system, handling refactoring, and doing reviews.
Today, however, an increasingly large part of these activities can be accelerated by AI tools.
This means that the value of the developer is shifting from "doing" to "deciding well what should be done."
One of the most important points in Anthropic's article is that AI becomes very strong when the goal is clear, especially on well-defined tasks.
But are we, as developers, still truly understanding what is being produced?
Because if AI generates code that works, but we can no longer explain it, maintain it, or assess its risks, then we have not really increased our productivity. We have only pushed technical debt a little further down the road.
In many teams, the historical problem was producing enough code. But with today's AI tools, the bottleneck shifts: it becomes the human ability to review, validate, and understand.
If a tool can generate in a few minutes what used to take hours, the critical point becomes how quickly we can understand, verify, and maintain it. This is where the developer of the near future is defined: someone who owns judgment skills that cannot be delegated to AI.
And so my provocation is this:
What if part of the low productivity of those teams does not depend on the tools, but on the fact that they keep working as if we were still in 2015?
Today we have tools that can accelerate coding, debugging, testing, review, documentation, and learning. But if we use them with old mindsets, rigid processes, and poorly collaborative habits, the gain remains limited.