Mainframe computers running code dating back to the 1960s are still vital to some banks, airlines and government departments, but the skilled engineers familiar with their COBOL computer language are mostly dead or retired. Now researchers say AI may be able to fill this skills gap and help maintain or replace these antiquated yet essential systems.
COBOL dates back to 1959 and was designed specifically for large, centralised mainframe computers, which carry out bulk data processing for large organisations. When these machines fell out of favour in recent decades and were replaced with many smaller servers, or even cloud services, the language was no longer taught at most universities and we began to rapidly lose expertise.
But many of those computer programs remain in use today, often composed of millions of lines of so-called undocumented code, meaning it lacks any explanatory notes. When the programs need updating, fixing or replacing, it can be hard to find competent engineers, says at FPT Software AI Center in Hanoi, Vietnam. 鈥淭his is a very serious problem,鈥� he says.
Advertisement
Bui and his colleagues have been training an AI model called XMainframe on COBOL code in the hope that it will be able to interpret code and rewrite it in other programming languages if necessary. In tests, they compared XMainframe鈥檚 performance on accurately summarising the purpose of COBOL code against several other AI models, including GPT 3.5 and GPT 4, two versions of the large language model behind ChatGPT. XMainframe鈥檚 score was six times higher than GPT 3.5, which was the most capable of the other models tested.
Other companies such as and are also working on AI models to write COBOL. AI won鈥檛 be able to replace the world鈥檚 COBOL code on its own, says Bui, but it can accelerate the work of the human coders we do have, helping them quickly understand massive and undocumented systems so that they can migrate them to newer technology. But the process may still take many years, he says.
The US-based recruitment firm , founded by Bill Hinshaw, places hundreds of programmers into short-term contracts annually, with clients ranging from banks to insurers and even sensitive government departments. Hinshaw is sceptical that AI can reliably and autonomously do what his experienced human coders can, or that companies would feel comfortable running AI-generated code on vast mainframes where the stakes are high. 鈥淎I will support people; people will never go away,鈥� he says.
He also has no concerns about the supply of human programmers for roles, although the youngest coders on his books are in their 40s and the oldest already well past pension age. Hinshaw is 82 and personally wrote some of the earliest COBOL code to control cash machines, which still runs today.
鈥淚鈥檝e been hearing about a shortage for over 25 years, of people not being available, getting old, passing away,鈥� he says. 鈥淢ost of [the contractors] are in the 60s, early 70s. They don鈥檛 want to quit; they want to keep working. So we don鈥檛 see a problem with COBOL now. Ten years from now, we may see a problem, but right now we don鈥檛 see a problem.鈥�
arXiv
