Technical · Four years in, and the part that surprised me
Building my own tools
I went in early on generative AI and went in wrong — two months of enthusiastic slop before I stopped. What changed it was building software instead of generating images.
The wrong start
When image generation took off I threw myself at it — ComfyUI, automated sites, accounts, anything I could generate. For about two months I made slop, enthusiastically. Then it turned. I was copying, not making, and the world worked that out about as fast as I did.
On the strength of those two months I nearly wrote off the whole field.
What changed it
Local models, LoRA training, then heavy workflow automation, and eventually agent-assisted coding. Within a few weeks of the last of those I was building software rather than generating pictures.
The difference wasn't the output. It was that my judgment was driving. I've spent a career watching tools get built by engineers for engineers, and watching artists file design notes that never got iterated on. This was the first time I had authority over how something worked and how it felt, and could act on it the same day.
The view I hold now
The bottleneck isn't the model. It's that the two groups now able to build tools are each missing the other half of the skillset, and neither is missing the half people assume.
That's a training problem rather than a tooling problem — which is the one kind of problem I already know how to solve.