Waters Wrap: The end of Cobol?

About a decade ago, I was having a drink with an engineer at a large bank headquartered in Asia, though he was working out of its Manhattan office. Actually, at this point, he had left the bank about a year earlier, but he kept on working as a consultant. 

The reason he was still collecting a paycheck was that he had built an internal auditing application that he—and only he—knew how to code, patch and update. Because his services were still very much needed, he ended up making more as a freelancer (hourly) than he had as a full-time employee. He then used that freelance money to build his own startup, which is still in business, though it has nothing to do with finance. 

I remember this conversation from time to time because he told me that he explicitly made the code for the app as complex as possible because he knew he wasn’t going to be long for the world of banking, but wanted to see if he couldn’t “engineer” himself a consulting gig. Ethics aside, his plan worked to perfection. 

It’s true that using generative AI for coding comes with risks, but the bigger risk lies in not adopting it

Operational risk manager at a G-Sib

On January 3, we published a story that examined how banks are experimenting with generative AI for coding. For the article, Menghan Xiao spoke with more than 20 sources and found that four global systemically important banks (G-Sibs) have begun testing to see how well it can be used to understand and translate legacy or proprietary code. It also explains how these bank technologists are exploring ways for genAI to be used in code improvement and generation. 

For the first use-case, sources told Menghan that we could see real-world offerings inside banks as soon as this year. For the second use-case, sources said 2025 (or later) was more likely.

“It’s true that using generative AI for coding comes with risks, but the bigger risk lies in not adopting it,” said an operational risk manager at a G-Sib.

Recently, I was speaking with the head of engineering at a large hedge fund, and I thought his opinions were interesting because hedge funds are not as highly regulated as G-Sibs. He also saw code translation as a prime area for genAI (and the large language models, or LLMs, that underpin these tools) to address in these early days. 

Look at it this way: If “X” in the English language means “Y” in French, and if you do nothing but make a big codex of all X–Y mappings and provided grammar rules, a model could go through a French newspaper and understand most of it—it might not be colloquially correct, but the reader will get the gist. The same can be done for code, said the engineering head. 

Additionally, genAI can help with semantic mapping, which has been a massive barrier for code modernization efforts. For example, there might be redundancy built in for a specific reason; maybe there are bug fixes unaccounted for; maybe there are nuances in the languages used or third-party libraries that aren’t clear from the code; or maybe there are things done for reasons that are unclear from the logic. What the model can do is help guide the user toward what’s syntactically and semantically correct from a programming perspective. 

“Essentially, what was the ‘intent’ behind the code? You can use machine learning to extrapolate that intent,” said the engineering head. 

But for that second piece—code generation—he said it’s still very early days and very experimental. 

“At banks where there’s tons of red tape, everyone is being very careful about putting new code into live production because everyone is fearful of being the next Knight Capital (which shut down in 2012 after losing millions due to a coding error). I’ve heard windows of like 2–10 years, where ‘2’ is being very optimistic where every developer will use AI and some kind of generative modeling to write any of their code. I’m not in the 2-year bucket, but I think 5 or so seems more reasonable.”

Good riddance?

So let’s end by looking at the primary use-case, code translation. 

Reading Menghan’s story reminded me of a story from April 2020 written by Rebecca Natale. It started with this quote: “LONG LIVE COBOL! It keeps me employed at 65 years old.”

Well, I hope he has been saving for retirement. I was recently speaking with a machine learning engineer at one of the largest trading technology companies in the world, and he believes it’s actually easier to map an LLM for Cobol than for more modern languages.

I was recently speaking with a machine learning engineer at one of the largest trading technology companies in the world, and he believes it’s actually easier to map a LLM for Cobol than for more modern languages

“The problem with Cobol—but the beauty of it for this particular task—is that it has constraints; you can’t do as many things as you can do with, say, Python.”

As a journalist, what scares me about the future of writing is how easily a chatbot like ChatGPT can create fiction. If you were so inclined, you could type: “Write me an idea for a story that is based off of the Hobbit but that takes place during the Victorian Era.” In seconds, you’ll have a treatment. Then you say, “OK, but add a humorous twist to it.” Sure enough, you’ll get a story about “Bilberry Baggins” trying to unlock the secrets of the “Chrono Mechanism” and, in the process, he “becomes an unintentional hero, navigating a clockwork jamboree.”

To simplify it, if you take the LLM and you train it on “here’s how Pythonic code would look, now here’s some Cobol language and make it look like that," it will get it into a certain shape where a developer can go through and manually fix anomalies or put comments in there like “blurry logic here, needs review.” You can also attach a confidence level to the output, says the vendor ML engineer.

“These language models can figure out when you say,  ‘Take my Cobol and turn it into Python,’ that you can do it in a way that isn’t just rewriting it line for line in Python just so that it works—you can actually rewrite it in a way that actually seems Pythonic in terms of using appropriate constructs that a reasonably skilled Python developer would use, as opposed to just raw-mapping the logic.”

For the capital markets, if 2023 was the year of the advent of genAI, then 2024 will be the year of rollout for low-hanging fruit … the “test” generation. Code translation, it would seem, is the most logical starting part for exploration. 

Got some interesting use-cases of your own? Hit me up: anthony.malakian@infopro-digital.com.

The image accompanying this column is “Church Street El” by Charles Sheeler, courtesy of the Cleveland Museum of Art’s open-access program.