Google engineers used big language models (AI) to speed up code migration within the company and achieved impressive results. A published study explains how the use of AI reduces the time spent moving code between different environments by 89%. Code migration is the process of rewriting from one programming language to another. about this reports edition of The Register.
The project involved several complex migrations: switching identifiers in the Google Ads codebase from 32-bit to 64-bit (more than 500 million lines of code), updating the testing library from JUnit3 to JUnit4, and replacing the Joda Time library with new code. standard Java package `java.time`.
Without the use of AI, porting 32-bit identifiers would require hundreds of man-years of work and complex cross-team coordination. Google’s AI-powered tools have automated much of the process. The engineer identified the identifiers that needed to be changed, and then the AI generated the updated code that passed the unit tests. The code was checked by an engineer and sent for review.
Even though 80% of the changes were created by AI, manual review and adjustment were required. In the case of the JUnit3-JUnit4 migration, 149,000 lines of code in 5,359 files were changed over three months, while 87% of the AI-generated code required no changes. The most significant increase in productivity (89%) was seen when switching from Joda Time to “java.time”.
The study authors emphasize that AIs are complementary to traditional traversal methods such as abstract syntax tree analysis (AST) and pattern searching (grep). While using AI is costly, Google has seen a significant reduction in migration time (50% on average) and a significant increase in the percentage of code written with AI.
Previously it became clear About a possible merger between TikTok and a US artificial intelligence company.