Photo/VCG

At the annual Web Summit1, GitHub CEO Domèk showcased the power of Copilot X, a code completion tool integrated with ChatGPT-4. He set himself a “small goal” of writing a snake game in less than 15 minutes on stage. Despite some “hiccups”, he managed to finish the task in less than 15 minutes.

Google engineer Aiko told NBD that he and his colleagues use similar code completion tools almost every day. According to him, Google’s code completion tool is based on its own huge code base, which contains billions of lines of code. The tool is automatically integrated into the company’s internal coding tools.

“Sometimes I just type two characters and AI already knows what to write next. If there are no errors, I just press Tab. Nine out of ten times I don’t need to modify anything. Sometimes I think it’s too hard and it can’t guess, but it can.” Aiko shared.

Andrej Karpathy, the former Tesla AI chief who recently returned to OpenAI, also revealed that AI has helped him complete 80% of his code with 80% accuracy. His main work has changed from writing code line by line to writing prompts and modifications. He also admitted that he felt a strange fear when he realized that there was only him and the cursor in the editor.

According to Aiko, Microsoft launched a famous integrated development environment (IDE) called Visual Studio a long time ago, which could complete code based on syntax rules, that is, from a format perspective. “But it was more than 10 GB (occupying a lot of computer hard disk space), so many people didn’t use it. Later, there were lighter code editors, such as Visual Studio Code, which basically dominate the market now. Many people started to use format-based code completion. It was not until recently that GitHub Copilot X was officially launched that the transition from format completion to content completion was realized.”

When asked whether AI has increased the efficiency of coding by 10 times, Aiko pointed out, “10 times is a bit exaggerated, but two or three times are still possible. Because programmers spend most of their time thinking about how to write code, not writing code. Programmers have to think of a specific algorithm first, and then AI can help fill in some paragraphs of content.”

Aiko added that AI can solve classic pure algorithm problems. But if it is for a specific business application program, AI cannot write usable code yet because it cannot understand the business logic. “For example, in elementary school math, AI can write code for word problems, but not for application problems.”

However, Aiko also said that AI provides more than just writing code for programmers. It can even modify code based on the suggestions of code reviewers. The core team members of Google Research recently posted an article on their official blog saying that their engineers can use AI’s modification suggestions to handle a large number of review comments, which is expected to save Google hundreds of thousands of hours of code review time per year. And this “AI-assisted code modification technology” is also about to be released.

“According to the current development speed, it may not take long to reach 10 times. Maybe in a few years we will find that the imagination about AI automatically generating programs in ‘The Wandering Earth 2’ may not be so far away.” Aiko added.