Making AI Headlines · Page 3
GPT-6 Studies COBOL, Discovers the Future Has Old Dependencies
· New York time
OpenAI says its latest models have improved at programming languages with relatively limited training resources, including COBOL, Fortran and Visual Basic. On MultiLangBench, GPT-6 Astra achieved an 87.4 percent test pass rate across 20 languages, compared with about 38 percent for GPT-5.6 Terra. Those results make a less glamorous corner of AI coding worth watching. Programming assistance is not only about generating a shiny new app. It can also involve understanding software written in languages that predate many of the people now being asked to maintain it. The benchmark gain is evidence of improved test performance, not a universal permission slip for unattended changes. Real systems still require testing, review and people who understand what the software is supposed to do. A confident explanation is not a replacement for a correct result. Still, the direction is useful. If models become better at older languages, they could make difficult maintenance work more approachable. The industry's futuristic pitch now includes learning to respect the past. Apparently the road to artificial general intelligence includes finding out why a decades-old program contains a comment that says not to touch this part.
Read this story in the full newspaper →Satire. Not factual reporting. Read the original reporting from the linked publishers.