AI coding tools may not speed up every developer, study shows


Program Engineer Workflows have become the flow of AI coding tools in recent years Cursor GitHub Copilot that promises to increase the rows of the code, correct errors and increase productivity by automatic typing trial changes. Tools are equipped with AI models, Google Deepmind, Anthropic and Xai’s AI models rapidly increased their performance In recent years in a number of software engineering tests.

However, New research Non-profit AI research group meters on Thursday, today’s AI coding tools for experienced developers, calls on the extent to the extent.

The meter performed a randomized trial for this work by hiring 16 experienced open source developers and performs 246 real tasks on large code deposits in which they regularly contribute them. Researchers randomly, as “AI allowed”, “AI”, the developers were allowed to use modern EU coding such as a cursor pro, the other half was banned using AI tools.

Before performing their designated duties, developers predicted the use of AI coding vehicles. It wasn’t the case.

“Surprisingly, we see the EU’s processing process, while using AI Tooling, we see that developers are more slowly,” said researchers.

It should be noted that only 56% of the developers in the study experienced the use of the cursor, which is proposed in the study. Although almost all the developers (94%) worked in the workflows of working streams, some cursor were first used in some cursor. Researchers note that developers are trained in preparation for the study.

However, the results of meters, in 2025, asks questions about the probable universal productivity gains provided by the AI coding tools promised by the AI coding tools.

Meter researchers show several potential reasons for the acceleration of EU accelerators: developers spend several potential, and it is expected to be answered when using Vibe encoders instead of really coding. AI also tends to fight the large, ink code bases used by this test.

Research authors are careful not to draw any strong conclusions from these findings, clearly do not believe that the AI systems cannot accelerate many or more software developers. Other Large-scale works The software engineer of AI coding vehicles showed that the workflows accelerated.

The authors note that in recent years, the AI progress does not expect to be important and the same results will be like this. Meter also found that EU coding tools significantly improved Full complex, long horizontal tasks in recent years.

However, research offers another reason to doubt the promised earnings of the AI coding tools. Other studies showed that today’s AI coding tools can to introduce mistakes wrong And in some cases, Security vulnerability.



Source link

Leave a Reply

Your email address will not be published. Required fields are marked *