Artificial intelligence: coders no longer live without

Artificial intelligence coders no longer live without

Sometimes criticized for their errors and their hallucinations, generative artificial intelligences (AI) have the particularity of excelling in computer code. Since its release last fall, a tool like ChatGPT very easily switches from English or French to Java, Python, PHP, C ++ or even Swift, these computer languages ​​which usually require long months of learning. Just write and the machine runs.

A simplicity well summed up by Andrej Karpathy, a former director of AI at OpenAI: “The hottest new programming language right now is English.” Dozens of tutorials on YouTube promise beginners to guide them in programming a website or an application from quick (commands) on the world’s most famous chatbot. “The same thing happened when technology made publishing easier. Or when digital tools made it easier to create music, movies, games. In all of these cases, we’ve seen an explosion of content and content creators,” Hadi Partovi, head of the Code.org organization, recently predicted on Twitter. Will the next tech nuggets be launched by strangers using AI-generated code?

Bargain for the pros

The OpenAI toy, in the first place, is not omniscient and errors persist. “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 arise”, explains the director of engineering of machine learning within the company Trail of Bits, Heidy Khlaaf, co-author of a scientific article focusing on the technical limits of the tool. “You still need prerequisites before getting started. Knowing how to recognize ‘clean’ code before you can deliver one thanks to AI”, confirms Alexandre Soyer, freelance web developer, author of the La Console newsletter.

At first, therefore, the wave mainly affects professionals. Besides ChatGPT, dozens of tools, like AlphaCode (Google), CodeWhisperer (Amazon), Cursor, or Copilot (Microsoft), are very popular among web developers and other engineers in their work. Everyone uses them on a daily basis or at least tries them”, confirms Frédéric Bardeau, director of the network of computer schools Simplon. Designed by the popular place of deposit and sharing of codes GitHub (Microsoft) and published several months before ChatGPT, Copilot is one of the most successful tools. It exceeded one million users at the beginning of the year. He’s not the only one to think so. Matt Welsh, a computer engineer who worked for Google and Apple, predicts that “all programs in the future will be written by AIs. These AI systems will fly our planes, run our power grids, and perhaps even govern entire countries,” said Mark Zuckerberg’s short-lived professor at Harvard University in the journal. Communications of the ACM.

“Cooking with pre-cut ingredients”

In the meantime, several very concrete use cases stand out, always under careful human supervision. Generative AIs are able to find errors in code and then fix them. The machine itself completes and suggests the lines to be written. “It’s like cooking with all the pre-cut ingredients”, remarks Alexandre Soyer. A ChatGPT, for example, can also explain a piece of code retrieved on the fly and finely detail its progress, like a real virtual assistant. Companies still dodge these tools for questions of security and confidentiality of their data.

But with a few protections and custom models, the lock could quickly break. “We can save a lot of time,” says the founder of Simplon. A recent study run by Copilot reports that, on average, its users accept nearly 30% of code suggestions. Alexandre Soyer arrives at the same percentage of time saved compared to his old standards. “This allows you to have more time for what cannot be automated, such as responding to a customer, listening to his needs…”, hopes Frédéric Bardeau, who has consequently upset the training and evaluation methods within his schools, taking into account these amazing AI capabilities. On the program: more critical thinking, role-playing with complex problems. A way to raise the level… At least until the next ChatGPT updates.

lep-life-health-03