AI is so effective that it could save companies money. (Image: azlen – adobe.stock.com)
Artificial intelligence is having more and more influence on our everyday working life. Researchers took this scenario to the extreme in a study.
A team of researchers from Brown University in the USA carried out an experiment in collaboration with several Chinese universities: They filled all important positions in a fictitious tech company with AI and had them program software.
The result of their work is quite impressive.
Those were the general conditions
The researchers created a hypothetical company called ChatDev and staffed all major departments using the ChatGPT 3.5 model.
A well-known model from the software industry was used. The waterfall model is used in software development. In cascades, the results fall one step lower until the program is finally finished.
The waterfall model is a well-known model in software development. (Image: Wikipedia)
In the case of testing, there were four departments: design, coding, testing and documentation.
The AI was given elementary roles and provided with various prompts. There was not one bot per department, but rather several, which were then divided into different hierarchies. For example, ChatDev’s CEO and CTO focused on the design phase.
Minimal human intervention
The AI decided everything itself. Through chat communication with each other, ChatDev employees have figured out everything themselves: from deciding which programming language to use to identifying errors in the code – until the software is finally ready. This also helped the researchers understand how the AI works and what path it follows.
The fictional company’s first task was to create a game of Gomoku (aka
Five in a row) to design.
You’ve probably already seen the game with the black and white stones, for example as a video game on the Switch. (Image: Nintendo)
In the first step, the AI discussed with itself which programming language would be used. The choice ultimately fell on Python because of its nature
Simplicity and readability, popular with both beginners and experienced developers.
After 70 different projects there was a clear cut. The AI took an average of seven minutes for a software and generated a total cost of less than one euro – and that with an included test phase.
According to the study, 86.66 percent of all software generated by ChatDev worked properly.
I tried it out myself that ChatGPT can’t do everything – with sausages.
Researchers give the all-clear
Even if the results of this study are impressive, the teams give the all-clear. Researchers have identified limitations such as errors and biases in linguistic models that could cause problems when creating software.
The role of a human developer and programmer is also too important. In addition, the software developed was all quite simple to use.
The researchers are convinced that the results of the study can still help aspiring programmers in the real world.
A team of researchers at Stanford University conducted a similar project – using Stardew Valley as inspiration. This example shows that ChatGPT doesn’t always get it right: A school library in Iowa uses AI to decide what flies off the shelves.
Researchers from several universities have started a series of projects and let AI run its own software company. The result is quite impressive. Do you think the study can be transferred to the real world? Could we at some point work hand in hand with AI in everyday life? Write your opinion in the comments.