“Let the floodgates of nerd culture open wide.” This is one of the slogans behind the new GitHub tool presented on Tuesday October 29, “Spark”, the spark in French. The principle: being able to bring any application or program idea to life using generative artificial intelligence. To manage a karaoke evening, for example, or to generate QR codes, lists Thomas Dohmke, the boss of GitHub, on the stage of the “Universe” event, organized last week by the company from the former military barracks from Fort Mason, San Francisco. Just run the “prompt” – the command – and Spark runs. Dohmke tries out the idea of a tic-tac-toe game with, instead of circles and crosses, ducks and hippos. Nothing very useful, of course, but the result has the merit of being displayed immediately (Dohmke wins his game live). The code, “hieroglyphs” for the uninitiated, is invisible to the presentation. GitHub, a software hosting and management service, hopes to encourage vocations: it is targeting a billion developers on its platform within several years, or ten times more than today.
One in eight humans, the bet may seem daring. At Fort Mason, an engineer from the company bought by Microsoft in 2018, says he is somewhat “surprised” by this little-publicized project internally. Among the developers present, some are downright doubtful. “Good luck installing these programs…” whispers a colleague, emphasizing the difficulty of going from idea to implementation; it’s a job. Security or bug issues can easily arise. But the democratization of this type of tool is, ultimately, not surprising, given the way in which they have established themselves among professionals. Nine out of ten developers use AI to code, according to GitHub, whose professional solution, Copilot, is the market leader. A year ago, a third of the Fortune 500 – the most successful American companies – used it, or a competitor (Cursor, CodeWhisperer, TabbyML…). “It has become almost obligatory to use it every day in order to remain competitive,” says Anan Kulkarni, CEO of Crowdbotics, met at Universe, this high mass of developers which celebrated its 10th anniversary this year. Particularly in the race for AI. It’s no coincidence that Python, the most widely used programming language for coding artificial intelligence applications, has recently become the most popular on GitHub.
“Doing without it has become a risk”
The biggest companies are already hooked. “More than a quarter of all new code at Google is generated by AI and then reviewed and accepted by engineers,” Sundar Pichai, its chief executive, said last month. Amazon, for its part, announced that it was saving $260 million thanks to generative AI applied to code. At Universe, the SAP software specialist provided feedback. Out of 23.6 million code suggestions made by AI, some 6 million are validated (25%). Individually, KPMG, on 1,500 licenses, cites 5.7 hours of time savings per employee per week. Out of ten of its developers, eight consider themselves more productive.
A feeling that many professionals encountered at Fort Mason share, without always putting statistics on it. “It helps me look at my work from other perspectives,” says Dwane, working for Vonage, a telecommunications app. “It’s like someone is always next to me to help me, sometimes to confront me.” Alex, manager at chip designer ARM, recalls that “what makes each developer unique is not the mastery of a particular computer language, but their ability to solve problems. this is what AI allows us to focus on.” “Giving up has become a risk,” adds Mark-Christian, developer at Postman. Something anachronistic, like a journalist typing his articles on a typewriter.
A KPMG manager says for his part that these tools are also an “argument” for recruiting developers, in the same way as advantageous health coverage, or teleworking days. “With the exception of a few holdouts, often very experienced and who feel they don’t need them, I don’t know anyone who would want to go back after testing these tools,” notes Colin, a tall blond coding at Docker. “That’s the goal,” says Jonathan Carter, of GitHub Next. “I’ve never had to measure whether the microwave was faster at reheating a dish than the oven. With generative AI for the code, it ‘is just as obvious.”
Developer deficit
In truth, everything is not so perfect. A study by an analysis company, Uplevel, noted no improvement in the productivity of AI-boosted coders. Worse, technology has introduced more than 40% more bugs into their work. Criticisms heard during the Universe event. “You should always be wary of technologies that promise to do everything for you,” considers Alex. “I don’t yet have confidence in AI,” says Forester, software architect at IBM. Dwane, from Vonage, has a motto: “Trust but verify”. The designers of these tools do not say the opposite. “Autonomous cars are not the best drivers. But they are better than the average driver,” Thomas Dohmke likes to point out. And that’s enough.
Because technology makes it possible to fill, in part, the deficit in specialized labor, to meet the growing needs for IT development, which is true in all sectors. An American public statistical agency estimated the lack of qualified workers at 40 million in 2020, but predicted that figure would double by 2030. This is particularly the case in the automotive sector, where daily updates of software that makes driving easier (and which even aim for autonomy) are now commonplace. At L’Express, the boss of GitHub also talks about his visit last year to Société Générale in Paris: “It’s a bank, but it requires the cloud, AI, software, without which it would put you behind.” There is an urgent need to code more and more, and faster. This is what generative AI tools have enabled for more than three years now. Codex, the first version of Copilot, although based on the same model as ChatGPT (GPT-3), was released earlier, in mid-2021. If it was content to sometimes offer to complete code (autocompletion), the latest features cover a multitude of tasks: bug detection, test generation, restructuring, etc. In other words, the complete range of the modern developer. And at a lower cost: between $19 and $39 for a monthly subscription to generative coding AI.
Code, more predictable than language
The question remains, why is progress visible more quickly in the code than elsewhere? “Programming language is much more predictable than text,” explains German Thomas Dohmke prosaically. The result is therefore generally more reliable compared to language, where AI has not yet become essential. “The boom in quality models in 2024”, according to Dohmke, accentuated the gap. Its solution, Copilot, now offers the possibility of choosing between several LLMs according to its constraints, like those of Anthropic (Claude) or Google (Gemini), in addition to those of OpenAI (ChatGPT). Each having taken particular care of their “coding” sections. Not really for the sake of science. But because they constitute a promising segment, financially.
GitHub alone claims 300 million in annual recurring revenue and a user growth rate of 180%. If calculation costs remain high, the profitability horizon seems closer than elsewhere. Automation always calls for more… automation. From a security point of view, many observers believe that AI will be very useful in navigating the billions of lines of code that are now dated, sometimes obsolete and therefore vulnerable to cyberattacks. There is no doubt that companies will pay a high price for this service. Finally, “agents” are now being deployed in the software. Either automatons taking charge of all or part of tasks, the AI planning and carrying out various related actions. Always promising greater speed and efficiency, the human being placed here more as a team leader, a supervisor.
For all these reasons, Jonathan Carter believes that “developers are at the forefront of generative AI”. They discover “its full potential” before other sectors. Indeed, they are already seeing real effects from it. This results in faster execution and increased productivity. However, this growing dependence on technology could ultimately lead to a loss of human skills. In any case, it is perhaps no coincidence that the name “Copilot”, which initially only designated the GitHub code generator, was taken up by Microsoft to name its entire AI suite. The code seems, in its own way, to lead the way.
.