“Software is eating the world”, software is eating the world, said the American entrepreneur Marc Andreessen. The statement today does not seem particularly risky and I do not think anyone tries to contradict him, but the interesting thing is that he said it ten years ago , in 2011. Currently his observation is more like a prediction, since there is practically no essential section in daily life that does not use software: even in modern cars there are millions of lines of code.
Teaching Artificial Intelligence to program is the goal of the most recent initiative of the IBM research department. Project CodeNet is the name of the project, which consists of a database with more than half a billion lines of code in fifty programming languages, from the oldest, such as COBOL or FORTRAN, to the most recent C++ or Python.
“The language of machines is code and, in the same way that we have advanced a lot in natural language processing and in translators, we can teach machines to program,” said Darío Gil, IBM senior vice president and director from IBMResearch.
Context is the key
The great challenge of this project is that teaching code is not so different from teaching a language, and all language, be it that of a human civilization or that of a machine, is strongly marked by the context. IBM Research explains it very well in the article with which it presents the project:
“If it was easy and rule-based systems worked, then the first programming languages, like COBOL, would have been converted long ago. But the programming language has a context. The meaning of any statement is related to a context, so doing the translation, as with human languages, is complex and time-consuming”.
With current technology, 50 to 60 percent of rule-based languages can be translated automatically, leaving the rest to be manually translated by a human.
The goal of Project CodeNet, therefore, is to provide innovative algorithms that can understand the context of programming languages, similar to what we do with human languages. In short, it is about getting the machines to understand the code, not just to process it.
