He is asked for information, explanations, sometimes even film scripts, poems or songs. Answers to all our questions. What if ChatGPT was ultimately never as good as talking to us in its native language? According to a confidential Google report unveiled by the American media CNBC, the tool developed by its competitor OpenAI would be able to be hired within its own company as an engineer specialized in coding. More specifically, at an “L3” grade, paid around $230,000 a year in Silicon Valley.
Criticized for its limits, its errors, its biases when it comes to discussing in English or French, the conversational agent allows, on the other hand, to program a website in a few clicks. All for commercial or personal purposes. Senior civil servant and essayist Raphaël Doan recently confided to L’Express that he had created his page aiorart.com (dedicated to the encounter between AI, art and literature) thanks to his help. Tutorials abound on Youtube, where Internet users publish their tips for designing software, web extensions or even mobile applications. Just write and ChatGPT runs. Which makes Andrej Karpathy, former director of AI at OpenAi and Tesla, say: “The hottest new programming language right now is English.”
On paper, this development is enough to upset the web as we know it. “The same thing happened when technology made it easier to publish. Or when digital tools made it easier to create music. Or movies. Or games. In all of these cases, we’ve seen an explosion of content and content creators”, predicts Hadi Partovi, head of the organization Code.org, on Twitter. This swarming, if it occurs, would be a “tremendous turnaround”, underlines Louis Derrac, consultant, expert in digital culture, compared to a current Internet very centralized around a handful of large platforms. The promise is unmistakably similar to that of the new era of the Web, sometimes called Web3.
Democratize programming with AI
Without going that far, OpenAI claims the desire to democratize programming. The company led by Sam Altman offers the possibility “for everyone to do more with computers”, she writes on her website. And she wants to go even further. According to the American media Semafor, out of nearly 1,000 recruitments recently made by the company, about 400 were for the purpose of understanding computer programming languages (eg JavaScript, Python, PHP, Ruby, etc.). The idea: to produce ever more complex and correct code and to be able to explain its progress in the dialog boxes. OpenAI also has another direct translation tool, from everyday language (English, French, etc.) to code, called Codex.
The firm is not the only one to progress in this area. AlphaCode, created by Google, pursues the same goal: to make code easier to write for people who are interested in it. And they are many. Evidenced by the strong growth in France, the number of short courses in web development for students or unemployed in professional retraining. Or the success of companies like OpenClassrooms, also oriented around learning computer programming, leader in online training in Europe. It’s no secret: the digital world is constantly looking for new talent. In France alone, more than 230,000 developer or engineer profiles will be missing by 2027, according to the professional union Numeum. Hadi Partovi believes in starting a virtuous circle. “Generative AI can automatically generate code, which makes it easier to build software and amplifies the power of a software engineer. This will accelerate the creation (and demand) of software, and more people will become software engineers. “
The first concerned remain cautious. “We are not going to immediately advise our students to use AI, believes Boris Paillard, CEO of Le Wagon, an intensive teaching network specializing in coding. But there is, thanks to it, a new way of make teaching less academic. Today, beginners in code sometimes give up because of the complexity of the languages and because there is a significant step before producing something useful or really creative.” And contrary to popular belief, you don’t need to like math to get started. According to a study published in Nature in spring 2020 and led by American neuroscientist Chantal Prat, “the skills needed to learn a programming language (like Python or Java) are remarkably similar to those needed to learn a second natural language”. In short, if he managed to learn Spanish, the brain can completely understand a computer language like JavaScript thanks to ChatGPT.
The limits of generative AIs in code
Obviously, the OpenAI toy can’t do everything yet. “One of the challenges in generating code is that the user’s intent must be captured well enough for the precision of the generated code to be correct. However, although natural languages are very expressive, they are very imprecise and ambiguities almost always appear”, explains to L’Express the doctor of computer science Heidy Khlaaf, co-author of a scientific article on Codex focusing on its technical but also societal limits. Even minimal computer knowledge is still recommended before using ChatGPT. If only to know what code to ask him, in what language, and for what functions. “Developing software, for example, requires not a few bits of code, but reasoning and several hundred files to be built in different programming languages”, recalls Boris Paillard. The epidemic of deepfakes, ransomware and hacks of all kinds that are shaking the Internet is finally a reminder that AI may not be in everyone’s hands.
The emergence of generative AI should therefore, initially, mainly concern the pros. “AI will save coders a lot of time,” predicted L’Express, Thomas Dohmke, CEO of GitHub. This huge development platform acquired in 2018 by Microsoft has also launched Copilot, an AI that automatically suggests lines of code based on what developers enter. A way for them to get rid of the most daunting tasks to focus on those with added value, requiring the most human reflection.
Some believe, however, that other technologies will, in the short term, have more impact on web development, the creation of websites or software. First, the “nocode” trend: programs from which, without the need for additional coding, it is possible to develop hundreds of different projects. “AI is not an end in itself, but a means to achieve it”, judge Emmanuеl Straschnοv, head of Bubble, a start-up in view of the sector. There is a risk, however, of tomorrow relegating men to a simple supervisory role. And to no longer really master the language of machines. This is the theory of Matt Welsh, a computer engineer who worked for Google and Apple, short-lived professor of Mark Zuckerberg at Harvard, in the journal Communications of the ACM. “It seems quite obvious to me that, Of course, all programs in the future will eventually be written by AIs […] These AI systems will fly our planes, run our power grids, and perhaps even govern entire countries.”