A recent study claims that artificial intelligence in software engineering, such as OpenAI’s ChatGPT, has the potential to operate software businesses’ operations more quickly and cost-effectively with little human intervention.
Brown University and many Chinese academic institutions collaborated to perform an experiment testing the capacity of AI bots (based on ChatGPT’s 3.5 model) to traverse the software development process without prior training.
The research framework included a fake software corporation called “ChatDev” to offer a real-world application setting. This entity was built using the traditional waterfall paradigm. This procedure, which typically included design, coding, testing, and documentation, was delegated to various AI bots.
Researchers assigned separate “vital details” to bots, which included a split of duties, communication methods, end goals, and other factors.
The AIs with the titles “CEO” and “CTO” were in charge of the design stage, while those with the titles “programmer” and “art designer” were in charge of the coding phase. These AI bots then proceeded to meticulously attack each aspect of the software’s progress. Their discussions spanned from programming language selection to bug identification.
In one example, ChatDev was charged with designing the blueprint for a basic Gomoku game. The conversation began with the AI ‘CEO’ querying the ‘CTO’ for the best programming language, with the latter praising Python for its “simplicity and readability.”
ChatDev’s findings were impressive after 70 different tasks: the software lifecycle was completed in under seven minutes and cost less than $1. Surprisingly, the process was effective at detecting possible problems due to its inherent “memory” and “self-reflective” capacities. Approximately 86.66% of the software developed operated flawlessly.
The study highlighted the speed and cost-effectiveness of the automated software generation process using ChatDev. It’s worth noting, though, that the researchers were unavailable for additional comment at the time this piece was published.