Sparring with AI
I built my first app over forty years ago. Using BASIC. On a Radio Shack TRS-80. It was an app that displayed a Christmas tree with blinking lights. I somehow convinced the RS sales guy to let me play on their one computer. In return I'd build something they could put in the store window to help with holiday sales. I never even thought to ask for commission. My 13-year-old interests were myopically focused on sports, Asteroids, and computers.
I had zero forethought that the future of computing would play out the way it did. But I did think it'd be cool to become proficient enough at coding to build games.
Fast-forward many decades (and languages and operating systems and platforms) later. While Christmas is still my favorite holiday, I haven't written an xmas app since. But I have built a lot of other programs in the intervening years. Enough to have a healthy derision for using LLMs to code. After all, so my arrogant mind went, these agents were likely trained on my code (and others, of course).
Making cute limericks and humorously morphed images? Sure. Writing secure, full-stack code that businesses actually depend on. No way.
I kinda, mostly, sorta still feel that way. But my defenses are weakening. Here's why:
I recently released an employee and asset management app for a friend's company. No AI. Not a chance. I didn't trust that it'd handle the security concerns associated with PII and multi-tenant access.
But now I'm working on one of my pet projects and have decided to give AI an opportunity to make me a convert (or not). Part of the challenge is that I love coding. Vibe coding, to me, would be like an artist asking AI to paint a picture for them. It might do it, and it might do a very good job. But have we then lost the art? It carries none of the artist's fingerprints. No struggle, no risk, no self. So having an agent crank out code for me holds little appeal. But ... having one read mine to review, test, poke holes, and challenge my thinking ... that has proven to be quite interesting.
So far I've only tested with Claude's latest, Opus 4.8. And I'm super impressed. Asked to identify improvements in an algorithm for dealing with vector anisotropy over a large corpus of text and ... it was good. Very good. Much better than mine. Ouch. But ouch in a good way. That said, it's no oracle. I've watched it propose database schemas and platform architectures that were flatly wrong, confidently so.
Where's my head now? While I'm not quite ready to take on discipleship, there's no doubt that human/AI collaboration can be a very powerful multiplier.
As to an impending AGI singularity? Well, just as my younger self had no clue as to the long-term future of computing, this much older version is equally in the dark. But I suspect it'll be a fun ride ... at least up to a point.