AI has meant that so many apps have never been launched in such a short time. The problem is that almost no one is using them.
A new research from MIT monitored the activity of thousands of developers before and after adopting the use of AI agents such as Claude Code or Codex. In their conclusions they have detected a singular “funnel effect”: many begin to use AI to “tinker” with the codebut few projects end in final software releases. Not only that: many more apps arrive in the application stores, but the interest they generate is practically zero. Hotfix blocks everything. When you start programming with Claude Codeeven without knowing how to do it, usually experience that feeling of “being able to do anything”, but although there are projects that manage to evolve and reach a successful conclusion, many remain in an attempt: the manual processes of quality control, code review or deployment They become bottlenecks for these projects, often carried out by people without a technical profile and without knowledge of software engineering. Producing more is not selling more. The MIT study, led by one of its economics professors, Mert Demirer, reveals another striking conclusion: producing more does not mean selling more. Although the volume of mobile applications that comes to the App Store or Google Play Store has grown spectacularly due to the ease of programming with AI, real consumption has not moved. There are many more apps, but the same consumers (and buyers), who are not downloading, trying or certainly buying more than they did. App failure. The vast majority of these new software products created with these tools fail spectacularly when it comes to capturing a minimum audience. The efficiency when creating them is not directly linked to the real usefulness of these tools or with the value that the market demands of them. The code surpasses us. Linus Torvalds already said it when speaking of how he saw the world of AI: “AI will be a tool, and it will make people more productive. I think vibe coding is great for getting people to start programming. I think (the code it generates) is going to be horrible to maintain… so I don’t think programmers will go away. You’ll still want to have people who know how to maintain the output.” His comment was a good prediction of what we are seeing now: too much code being distributed everywhere makes a review of that code necessary to validate it, but doing something like that at the rate that AI is adding code to these projects is unaffordable for human programmers, who are being overwhelmed. And costs are skyrocketing. The new AI agents that program consume a large number of tokens, which has made set off the alarms in companies like Uber or Microsoft. To mitigate the problem, a hybrid usage model is now being imposed: AI agents in the cloud, such as Claude Opus 4.8, are used to plan development, but then the code is written by cheaper cloud models, which are remarkably compliant, or even local models if the user/company has sufficient local computing capacity. Remembering the industrial revolution. Demirer remembers that something similar happened during the industrial revolution. The first factories that replaced steam engines with giant electric motors did not lead to a large increase in productivity. The real boost to efficiency came decades later, when engineers redesigned those factories from scratch by installing individual motors at each workstation. In Xataka | We believed that the AI talent war is about engineers and developers. Actually, it’s about plumbers and electricians.