“This method has the potential to transform the best way code is maintained in massive enterprises,” a gaggle of authors from Google Core and Google Advertisements wrote in a brand new “expertise report” describing their method. “Not solely can it speed up the work of engineers, however make potential efforts that have been beforehand infeasible because of the enormous funding wanted.”
In the end, Google’s aim was to establish alternatives for LLMs to supply further worth and help scale with out requiring difficult-to-maintain summary syntax timber (AST). These are broadly used to symbolize the construction of a program or snippet of code, however they’re deterministic — that’s, outcomes are already recognized — and code migration use circumstances span far more complicated constructs which might be tough to symbolize with ASTs, the authors defined.
“Attaining success in LLM-based code migration shouldn’t be simple,” the authors famous. “Using LLMs alone via easy prompting shouldn’t be adequate for something however the easiest of migrations. As a substitute, as we discovered via our journeys, and as described within the case research on this paper, a mix of AST-based strategies, heuristics, and LLMs are wanted to attain success. Furthermore, rolling out the adjustments in a protected technique to keep away from expensive regressions can be essential.”
