I would ignore my normal; "It's been so long"; but... why break a trend. It's been a while. I'm not digging into anything that I need to blog my thoughts out.
Even this - no problem to work through or express. Just... What the fuck to do about GenAI?
I've been using Claude Code HEAVY since mid-2025. It's a tool. Gotta learn the new tool.
I rewrote my magic card side using claude entirely. I'm rewriting it again using claude.
My 25 years of software development; with an EXCESSIVE focus on code quality and design give me a foundation to use AI from. I'm learning nothing.
The new site uses React and Postgres; I know nothing of those. I have an entire website running on those technologies and... I know nothing.
COULD I?
Yes.
AM I?
No.
Since I've started with AI; I've seen one inevitable outcome; Technopriests. The words of invocation will be uttered and the system shall respond. The what produced will become irrelevant. Enough focused agents, enough forests burned; the code will be functional, and secure, enough. All we'll be left with is the chants used to invoke the magic.
Some of us will be the code-speakers. Those masochistic enough to wade the genAI madness to fix that which it doesn't yet understand.
I have a very particular way of writing code; and I enforce that on what the agents produce. I have, largely, abdicated the front end, and the database, to whatever the agent does. The C#, that's my realm and dear fucking god it's bad at it. So I have little hope for the other realms.
One driver to write this ... Mythos.
For months it was hyped as a god-tier LLM. I had it for those few brief... glorious... days... It's ... hyped... but holy fucking hell... that's the first LLM I felt was deserving of being capable of jr engineer level work. Other's have called it a Sr engineer; but I know sr's that have been coding for 4 years. Sooooo.... Sure? But no; it's a fucking Jr engineer.
The key is that it still doesn't THINK. It still does the wrong thing. The biggest difference for me with Fable was that it did what already existed. OPUS put in the wrong thing, it made it up; did it wrong. Fable copied that; IN THE FOLDER IT EXISTED. And did it right where the right example existed.
It didn't 'make up' new ways to do things. It followed the examples. This is where I see a turning point for the use of LLMs for code. It uses the examples more explicitly and intelligently. I'm back to OPUS and I have to correct the code ALL the time. Mythos... I had to correct OPUS's mistakes in the mythos code. That's it. That... was fucking scary.
As an industry; we will end up curating a swarm of agents. We'll have the human mind do a final sign off; not even a review. There will be such a high volume that looking at it all is gonna be impossible.
Let's take my simple, low volume, magic card site. I like my code to be patterned; "Make similar things more similar". This enables me to quickly scan the code. If I spend 10 minutes validating Class-A; and Class-B does, basically, the same thing, and has the same code shape - I don't have to dig deep. I trust that it's doing it the same way. The literal file looks the same, so I assume it's doing the same.
And don't ask, "why don't you create an abstraction"... I fucking have. ... It makes the code harder to read. It makes the code harder to change. It makes the code harder to validate.
... 'why haven't I created an abstraction'... You're reading this; you should know I create abstractions to the point of something turning in to a pure pass-through... and then rolling back one level.
I'm not starting in on creating RAGs or training my own LLM or running shit locally. I'm still exploring the quickly evolving tool set. And it's scary to see what this latest leap provided.
I'm gonna burn tokens and have Fable do a clean room implementation of my site. "Fable, create a spec for this site" -> "Fable, create this spec'd site". No guard rails, not markdowns, no nothing - What does it generate when just given requirements?
I need this because the one thing that genAI doesn't do is architecture. It has the words of architecture; but the words and the code are so separate that it won't have the training data showing it how to get to good architecture implementation. I want to see what Fable, "wants" to produce...
Nothing much... I've got AI thoughts bubbling around. I hope to splash out more of these periodically