Digital Event Horizon
Google has successfully harnessed the power of large language models (LLMs) to revolutionize code migration, significantly reducing the time required for internal migrations by up to 89 percent.
Google's innovative approach to code migration using large language models (LLMs) has been shown to reduce migration time by up to 89% compared to manual methods.The LLM-based workflow involved a multi-step process, including AI-generated verified changes and human review.About 80% of modifications in the change lists were purely the result of AI, highlighting its potential for automating code migration processes.Google successfully migrated 5,359 files and modified approximately 149,000 lines of code to transition from JUnit3 testing library to JUnit4 in just three months.The study emphasizes that LLMs should be viewed as complementary tools rather than a replacement for traditional migration techniques.
Google, a pioneer in artificial intelligence (AI) research and development, has made significant strides in harnessing the power of AI to accelerate code migration. In a groundbreaking study published in a pre-print paper, a team of Google computer scientists revealed that their innovative approach using large language models (LLMs) significantly reduced the time required for internal code migrations.
The researchers' focus was on bespoke AI tools developed for specific product areas, such as Ads, Search, Workspace, and YouTube. These specialized LLMs were designed to assist with code transformation tasks, which are essential for maintaining large codebases in enterprises. The study's findings demonstrate that this AI-driven approach can reduce the time required for migrations by up to 89 percent compared to manual methods.
According to the researchers, their LLM-based workflow involved a multi-step process. First, an engineer from Ads would identify specific IDs requiring migration using a combination of code search, Kythe, and custom scripts. Next, an LLM-based migration toolkit was triggered, generating verified changes containing code that passed unit tests. These changes were then manually reviewed by the same engineer or other team members. The final step involved sending the code changes to multiple reviewers responsible for the affected portion of the codebase.
The researchers observed that 80 percent of the modifications in the change lists (CLs) were purely the result of AI; the remainder consisted of either human-authored changes or human-edited suggestions. This highlights the significant potential of LLMs in automating and streamlining code migration processes.
In the case of Google's Ads product, the researchers migrated 5,359 files and modified approximately 149,000 lines of code to transition from JUnit3 testing library to JUnit4. This effort took just three months to complete, reducing the time required by an impressive 50 percent compared to manual methods.
While the study's findings are encouraging, it is essential to acknowledge that LLMs should be viewed as complementary tools rather than a replacement for traditional migration techniques. The researchers noted that additional tooling may be necessary to prevent the human review process from becoming a bottleneck and to address concerns about the cost-effectiveness of using AI in code migrations.
The researchers concluded that LLMs offer significant opportunities for assisting, modernizing, and updating large codebases. These tools come with flexibility, allowing various code transformation tasks to be framed within a similar workflow and achieved success. As such, this approach has the potential to profoundly change the way code is maintained in large enterprises.
In an era where AI and machine learning are increasingly being harnessed for software development, Google's innovative approach to code migration using LLMs serves as a shining example of how technology can be used to enhance productivity, reduce costs, and accelerate innovation. As we continue to explore the vast potential of AI in software development, it is crucial that researchers and developers collaborate to develop tools like those described in this study.
Related Information:
https://go.theregister.com/feed/www.theregister.com/2025/01/16/google_ai_code_migration/
Published: Thu Jan 16 12:22:59 2025 by llama3.2 3B Q4_K_M
