What’s scarier? Finding out that your checking and 401K are managed by decades old COBOL programming? Or that AI will replace it almost overnight?

Photo above - official headshot of Dario Amodei, CEO of Anthropic. He predicted in his blog yesterday (once again) that his Claude AI system will soon write all the software code on planet Earth. Is Dario the Elon Musk of AI?

Full disclosure: I used to work at a money center bank. (Call center operations). Very labor intensive. The only thing MORE expensive/labor intensive was the IT (Information Technology) department. Millions and millions of bucks annually to repair and sustain ancient COBOL software written nearly a generation ago.

COBOL is expensive to repair and maintain that some bank and wall street IT departments gave up. They offshored this job to outfits in India and Pakistan. This is cheaper, but those coders typically disappear forever 15 seconds after the job is done. They left behind minimal documentation about their work. Anyone who sat through an end-user requirements or test review meeting where half the participants are on teleconference from Karachi and Mumba knows the pain of this. (What the eff did he say again? No, turning the volume up DOESNT help. It makes it worse)

Anthropic announced it can do away with all that. Just turn the coding over to Claude AI. Software update cycles will shrink from years to a few weeks. Most of those offshore contractors never get hired. This seems like it SHOULD be a win-win, right?

People who own IBM stock apparently think so. In the midst of yesterday’s snow bomb cyclone there was a blizzard of Wall Street sell orders for IBM. Share prices dropped the most in 25 years within just a few hours. Because IBM consultants and programmers are apparently part of the secret sauce that keeps this ancient tuna and mac casserole edible. And Claude is going to fix everything.

And maybe it can. Maybe AI will actually be able to build viable defenses against all the programming flaws that bedevil money management. Hackers, ransomware, DDOS attacks, wallet thefts, identity spoofing, bogus orders. Maybe these will all be fixed by a couple of chips in a little black Anthropic box, and we can kick back and watch Netflix.

Except the little black cable box already under my TV can’t even get real time closed captioning right. AI can't legally or competently drive a car. The last time it tried (with Tesla) self driving software had 4X as many accidents as human drivers.

According to various surveys, It takes a dedicated team of human software engineers LONGER to finalize AI generated code than it would have to make it from scratch with just humans. But the silver lining is that there are fewer humans involved, from start to finish).

Let's imagine the future. Your nearby ATM doesn't work. The bank website is down. How long will it take human repairmen to take apart AI Cobol code and find out what’s happening? There may not be consultants or contractors who have ANY IDEA what the AI bots wrote, and which lines of code failed.

I don’t predict that AI will be a complete failure forever at writing code. Just that selling your IBM stock because of some Anthropic press release on a snow day is probably an overreaction.

I’m just sayin’ . . .

https://financialpost.com/technology/ibm-shares-plunge-anthropic-touts-cobol-modernization

DeWayfarer · 61-69, M

AI can't. There are quirks in COBOL that AI just can't understand. Heck AI has a problem with nuanced English even. It's the same problem.

GPT-4o mini

It's true that COBOL, like many programming languages, has its unique quirks and nuances that can be challenging, even for AI. The complexity sometimes arises from the language's extensive legacy and syntax rules, which can lead to misunderstandings in interpretation.

Understanding the Challenges
1. Syntax and Structure

COBOL's verbose and specific syntax can be difficult to parse accurately. Unlike many contemporary languages that offer more flexibility, COBOL's strict formatting rules present a steep learning curve.

2. Contextual Nuances

Just as AI can struggle with nuanced English due to context-dependent meanings and idiomatic expressions, COBOL can also have context-specific implementations that aren’t immediately clear.

3. Legacy Systems

Many COBOL applications interact with outdated systems, leading to challenges in integrating modern AI tools designed for newer languages and frameworks.

4. Interpretation of Ambiguities

Natural language often contains ambiguities that require human intuition to unravel. Similarly, in COBOL, the intended function of a snippet may not be immediately evident without contextual knowledge of the larger program.

And this is just a start. It doesn't cover linkages and the placement of those links. Modify the code you alter the linkages path.

To get around this problem you simply must be creative in your approach. AI lacks that creativity.

And it's a pain to document why you had to do it just a certain way, when you really don't understand why precisely. You just know that is the over all problem intuitively.

There's a reason why nobody wants to program in COBOL. 🤣 Repeated intuitive trail and error isn't very rewarding. 🫤