The time savings can be numerous, but be careful to understand the generation of the code to better intervene on it.
AI particularly affects developers. It can generate code, debug, or even detect anomalies. AI can also generate advice at each stage of a project, and adapt to the level of its interlocutor. This makes him, in theory, an ideal teacher for budding developers looking to improve their coding skills. But a methodology must be adopted.
ChatGPT takes the lion’s share of tools
To learn to code with AI, the tools used by junior developers can be grouped schematically into large families. Conversational code assistants, like Anthropic’s Claude and ChatGPT, make up the majority of those used. The OpenAI conversational module, known for its ease of access for beginners, is particularly requested by our interlocutors. Gemini is also used, in addition to ChatGPT, by Alexandre Chevalier, a computer science student. It is completed by Perplexity AI and DeepSeek from Adrien N’Kunga, junior fullStack developer. Thom Costuas, web developer at Inodia, recently replaced it with Mistral in his agency. “The answers differ little depending on the use we have of it, but the processing of the data remains a reassuring argument.” The goal is to receive clear and accessible answers to their questions.
The second category of tools is that of “hybrid code assistants with conversational capabilities”, such as GitHub Copilot, which affects apprentice coders Evann Hislers, journalist in training, specialist in AI and data. The tool’s ability to intervene directly in a project appealed to him. “As a student at School 42, a teacher-less, project-based programming school, I often have to navigate between several code files. Copilot can access all of these files, which allows it to be comprehensive. Often, I ask it to help me code something or to explain the best way to do it. It also helps me when I have errors and I can’t find them.” It also uses Gemini’s Canvas functionality. This allows him to create interactive mini-apps in his browser. “I especially need it when I want to understand a complicated concept, like algorithms for example, where it immediately becomes more visual with diagrams.”
10 to 30% more productivity
Typically, studies show that developers report a 10-30% increase in productivity when using AI. It can reduce repetitive steps, speed up tests and detect certain errors more efficiently. The coders surveyed also highlight the ability of AI to allow them to build projects that were previously inaccessible, in particular by shortening deadlines. “Thanks to these tools, I was able to design a mobile weather application with technologies learned with AI,” illustrates Adrien N’Kunga. Alexandre Chevalier, junior software engineer at Capgemini, explains for his part: “I used this process for a video game project that is close to my heart. I program it in basic C, without extensions like C++ or C#. It contains a library to manage the display and sound in particular. Certain specificities of the language were quite difficult for me to understand at the beginning. On the Internet, information could also be difficult to find. The AI helped me a lot, especially for function pointers.” Thom Costuas admits: “The AI in particular quickly reminded me of syntaxes on languages or technologies that I occasionally use, or found resources that search engines have difficulty bringing up. For example, I was able to retrieve API documentation with Mistral for an online marketplace platform whose name had changed. This saved me from having to request the resource from technical support.”
Knowing how to put AI aside
To achieve these results, methodological rigor is often adopted among the junior developers surveyed. They are aware that AI is, to use the expression, “a great servant but a terrible master”. It helps by easily generating code. But this ease can partly eliminate the upstream reflection and subsequent design phase. It is indeed tempting to delegate different phases of the work to him.
To avoid this, the junior developers interviewed admit to trying to put AI aside when a difficulty arises. To cope, they primarily use “classic” resources: official documentation, their courses, their own thinking, Internet research. “It is better to think for a few minutes about finding an error during compilation than to immediately ask the AI to solve the problem,” says Evann Hislers. “For example, this allows you to know where to start when you write code. This sometimes avoids having to completely recode certain things implemented because you haven’t thought about them before.” It is only when faced with an insurmountable obstacle that the developers interviewed can resort to AI. “I then ask him, a bit as if he were a teacher, to explain the concept to me and give examples,” points out Alexandre Chevalier.
Note that during this phase, the process of learning to code with AI can be planned from the start. By first defining the goal, for example: “I want to build a to-do list app with React.” Then, by asking the AI: “What are the essential React concepts I need to understand?”. Unknown concepts can then be worked on gradually.
Once the response has been generated, let us indicate that caution is required among the coders questioned. Overall, 46% of developers say they don’t trust the accuracy of AI output, up from 31% the year before, according to a recent Stack Overflow survey. Is this due to the proliferation of AI tools, with their sometimes impressive capabilities and often tantalizing feats? “I’m checking whether his solution is good practice or not,” replies Thom Costuas. Conversely, giving the AI a lot of freedom can lead to various difficulties. “For example, I created a PDF manager in one hour with Antigravity to split documents and merge others,” says Evann Hislers. “Even though I’m familiar with it, I haven’t looked at the code in depth. So there may be security holes.”
