Writing code that works on its own is no longer sufficient evidence of having learned programming, according to Ardit Sulce, a Python instructor whose courses have enrolled more than 650,000 students on Udemy. As AI tools become capable of producing large parts of applications, the emphasis shifts from writing every line to understanding what the tools produce, reviewing it, identifying its errors, and knowing how the system’s components connect to one another.
This view appeared in an interview published by JetBrains Blog with Sulce, who has 10 years of experience teaching programming and holds a master’s degree in geotechnologies from the University of Münster in Germany. He is also the founder and lead developer of ActiveSkill Dev, a platform that uses AI as a personal tutor for learning programming and software development.
AI Has Changed the Questions More Than It Has Eliminated the Teacher’s Role
Sulce says that the number of daily questions he receives from students has fallen to between five and 10 times fewer than before the spread of generative AI. When students encounter a problem, they can get an immediate answer instead of posting a question on the discussion page and waiting for the instructor or teaching assistant.
However, the questions that remain, as he describes them, have become broader and more dependent on judgment: Which path should be followed? Is the project suitable for inclusion in a portfolio? And how does what the student is learning fit into the broader picture? From this perspective, an educational course is not limited to answering a specific question; rather, it provides an organized path through a field in which beginners do not even know their fundamental questions.
Learners’ expectations have also changed. A Python course that ignores the language’s relationship with AI seems outdated to them; they want to learn how to use tools to accelerate code writing and build applications based on language models, agents, and automation. At the same time, the teacher remains important for determining what should be learned and in what order, and for providing the motivation and curiosity that Sulce believes AI tools still find difficult to reproduce automatically.
Not All Educational Struggles Are Equal
Sulce distinguishes between effort that builds understanding and friction that wastes time. Trying to determine why a loop executes one additional time, or tracing the flow of data through a program, can be part of learning. By contrast, problems involving the installation of a broken driver, an unclear configuration error, or an outdated lesson that is no longer compatible with the current environment are obstacles that can be delegated to AI without an obvious educational loss.
Under this rule, beginners should ask: Will solving this problem give me an understanding I can use later, or will it only return me to the point where I started? Sulce emphasizes that writing code manually remains important during the learning stage because it helps build a mental model of variables, loops, and data flow. In professional work, however, manually writing every piece of code may become less important than reviewing what the tools produce.
Three Skills for Developers at the Beginning of Their Careers
- Reading code they did not write: Developers should read GitHub repositories, their old projects, and agents’ outputs, and decide whether the code is correct and remains suitable when inputs change.
- Expressing requirements precisely: Sulce believes that a good request to a model is a specification defining what the code should do and what it should not do, as well as how it should behave in edge cases. This is close to the concept of requirements, which was often the responsibility of another party.
- Understanding systems at the system level: Knowing isolated pieces of code is not enough; developers need to understand the components of an application and how they are interconnected. This is the skill that makes them supervisors of AI rather than subordinates to its outputs.
This entails a change in how students and job candidates are evaluated. Instead of merely asking someone to build an application, Sulce suggests providing AI-generated code containing a precise problem, such as an error in loop boundaries, silent swallowing of exceptions, or incorrect behavior when an input is empty, and then testing the person’s ability to identify and explain the problem. A live debugging session can also be conducted to observe how the person narrows down the scope of the defect.
A Gradual Path from Code to Supervision
Sulce divides learning into three stages: manually writing real, small programs inside an integrated development environment; then using AI as a teacher to explain what the learner does not understand while moving on to larger projects; and finally using it as a work colleague that writes parts of the code while the learner reviews them.
This progression assumes that project size will increase gradually: from a script, to an application that uses a database, then a project that handles users, followed by a deployed service and an application that connects to an API. In practice, the path links knowledge of syntax with an understanding of systems, rather than assuming that architectural understanding will appear automatically after years of writing code.
The interview confirms that learning within a professional development environment is not a merely formal detail. AI-based projects include API keys in environment files, virtual environments, dependency conflicts, differences between package versions, and files created or modified by agents. Therefore, the ability to run the project, review changes, debug failures, and modify code safely is what determines whether the learner is genuinely supervising the tool or merely accepting what it produces.