
Dario Funds Middlemen. Elon Funds Kids.
October 19, 2026
Biblical Code for Owning Your AI Slave
October 22, 2026Listen to Free Audio Version Here
Somewhere right now, a bank is moving your paycheck with code that may be older than you are. Much of it is written in COBOL, a language born in 1959, and the people who wrote it are mostly retired, or dead. The world’s money runs on it anyway. Every so often a state unemployment system crashes or a bank’s mainframe hiccups, and somebody goes looking for a seventy-year-old programmer who still remembers how it works.
This week OpenAI says it may have found a replacement for the seventy-year-old.
Here’s what they published. According to OpenAI’s own evaluations, a version of GPT-6 called Astra scored 87.4 percent on something called MultiLangBench, a test across 20 lesser-used programming languages. The model before it, GPT-5.6 Terra, scored about 38 percent. That is not a small step. That is more than doubling. On a separate test that asks a model to migrate COBOL into a modern language, a version called GPT-6.1 Sol got 86.6 percent right when it was allowed to think only a little, and 97.9 percent when it was given more time to reason.
Now the caveat, and it’s a big one. These are OpenAI’s numbers, about OpenAI’s models, on tests OpenAI chose to publish. As far as I can tell, nobody independent has checked them yet. I’m not saying they’re wrong. I’m saying a company grading its own homework deserves a raised eyebrow until someone else looks at the paper. The next real test will come when other people run the same benchmarks and publish their own scores.
But suppose the numbers hold up even roughly. Why should anybody outside a server room care?
Because COBOL and Fortran are the plumbing of modern life. Banks, insurance companies, payroll systems, government benefit programs, airline reservations, and a lot of scientific and engineering software still sit on code written in the 1960s, 70s and 80s. Nobody rewrote it, because it worked, and because rewriting it was terrifying. When you touch code that moves trillions of dollars, one misplaced decimal is not a bug. It’s a headline. So for decades the strategy has been simple: don’t touch it, and pray the last guy who understands it doesn’t retire to Florida.
Well, he retired to Florida.
And here comes the gallows humor. For half a century the world’s financial system has been running on code that hardly anyone alive understands. The rescue plan is a machine that nobody fully understands either. We’re replacing a mystery with a bigger mystery, and calling it modernization. I’m not even against it. I just think we should be honest about what we’re doing. We are not removing the black box. We’re handing the old black box to a new black box and asking it to translate.
Think about what 97.9 percent means in that world. In a high school, 97.9 percent is an A. In a payroll system with ten million lines of code, 2.1 percent wrong can be a lot of wrong. So the honest version of this story is not “AI fixes the banks.” It’s “AI may finally make it cheap enough to start fixing the banks, with humans checking every line that matters.” That’s still a big deal. The hard part of these migrations has never been typing the new code. It’s understanding what the old code was really doing, including all the strange little exceptions somebody added in 1983 and never wrote down. If a machine can read that and explain it, the bottleneck moves from finding the last COBOL programmer to reviewing the machine’s work. That is a better problem to have.
There’s also a quiet irony here. For years we were told AI would come for the newest jobs first, the coders writing apps and websites. Instead, one of its most useful tricks may be reaching back in time and rescuing the oldest code on Earth. The future’s first big favor to us may be cleaning up after the past.
Now for some high-end pondering.
Every civilization runs on something it no longer fully understands. After Rome fell, Europe largely lost the knack for Roman concrete, and medieval villagers lived among aqueducts they couldn’t have built. We’re in the same place with our software. The systems that hold our savings, our pensions and our benefits were built by a generation that is leaving, and the knowledge left with them. Now we’re handing that inheritance to a mind of our own making, one that learned from everything that generation wrote down. In a strange way, the old programmers are still talking. Their comments, their manuals and their forum posts from 1997 are in the training data. The machine is reading what they left behind, the way a historian reads letters from the dead, and then speaking for them.
So who understands the system now? Not the original authors. Not the new engineers. Maybe the machine does, or maybe it’s just very good at sounding like it does. That’s the question I’d want answered before I let it touch my retirement account. I suspect we’ll find out the way we always do: one migration at a time, with a lot of very nervous people watching the decimals.
Until then, if you know a retired COBOL programmer, buy him lunch. He may be the last human who can check the machine’s homework.
Sources:
My Living AI News Now. Your Place for AI News. Commentary, not financial advice.
C. Rich



