Researchers at Brown University in the US and several universities in China conducted experiments to see if the latest version 3.5 of ChatGPT could complete the software development process without any training.
According to the research paper, a fictitious software development company called 'ChatDev' was first created for this test. The company is then divided into four phases – designing, coding, testing and documenting – based on the waterfall model (a hierarchical approach to developing software).
At this stage the AI bot is given instructions along with 'important details' to perform specific roles. All important topics like 'defined tasks and roles', 'communication protocols', 'work completion criteria' and 'limitations' are included in the description.
Following the allocation of distinct roles to the AI bots, each bot actively participates in its designated task. For instance, the 'CEO' and 'CTO' operate at the 'Design' tier of Chatdev, while the 'Programmer' and 'Art Designer' function at the 'Coding' tier. AI operatives within each tier engage in collaborative conversations and autonomously execute specific segments of the software development process with minimal human intervention. This AI bot takes charge of all decision-making aspects, ranging from the choice of programming language to software development, including the detection of internal code glitches.
Researchers examine software development in various contexts. The time and cost of each type of software development is analysed.
Chatdev was instructed by the researchers to design a basic 'Gomoku' board game. Evidently, after entrusting ChatDev with 70 tasks, each software development process was efficiently accomplished within a mere 7-minute timeframe, incurring a cost of less than a single dollar.