AI shortens programming from four weeks to 7 minutes

ChatDev, a team of Chinese and American engineers, said it has made new progress in applying artificial intelligence to programming. Instead of using a single AI, engineers gather multiple AIs into groups, helping to speed up the coding process.

Specifically, ChatDev develops AIs based on the ChatGPT 3.5 model, then sets them up to work together like members of a technology company. Each AI takes on its own role, hierarchically from large to small, including chief executive officer (CEO), chief technology officer (CTO), product director (CPO), graphic designer, and auditor. code review, programmer. Low-level AI must listen to orders from higher-level AI.

“We just make requests. The AI team will discuss on their own to find solutions. Finally, they will agree on the programming language, software interface, output results, and make corrections if necessary. need,” ChatDev representative said.

Engineers discovered many interesting points when observing the activities of the artificial intelligence team. For example, AIs have agreed to exclude CEOs from technical issues. In other words, the “virtual” CEO only gives initial information and provides summary feedback. AI programmers and designers can safely do their work without much impact.

“The CEO’s lack of deep intervention will avoid conflicts within the group. This is completely different from our real world,” said a ChatDev representative. The AI team also holds discussions to self-evaluate each step of the product.

ChatDev assigned 70 different tasks to the AI team. As a result, more than 86% of tasks were performed well, the rest encountered errors due to limitations in ChatGPT’s API. Each mission takes an average of 6 minutes and 49 seconds. Meanwhile, if replaced by humans, the process of completing work with similar complexity would take at least four weeks, or even months.

According to engineers, the test did not fully demonstrate the power of AI because it used an old version of ChatGPT. Besides saving time, using AI in groups also helps reduce costs, because the price of each completed software only fluctuates around one USD.

The test results are raising concerns that AI will take over programmers’ jobs. However, ChatDev engineers say the transition is not immediate.

“AI still needs precise settings from humans to operate. A small deviation can significantly change the output results and make products created by AI unusable,” said a ChatDev representative. “We will now see a rapid increase in human-AI collaboration.”

Hoang Giang