A technical credential and actual technical fluency aren't the same thing. I learned that working on an agent that kept getting information wrong.
Everyone's first instinct was a bug. A code error, maybe something in the platform. That's the reasonable place to look, and it's where a technical background trains you to look.
The problem was in how the information had been structured. What it was being asked to do, and in what order. How it decided which source to trust. What it assumed when a request was vague. So I went in and rebuilt the reasoning sitting on top of the model, and the information came back clean. Nobody needed a better model or a new library.
I've always been drawn to tech even though I come from a non-technical background, and what building agents and integrations taught me is that the work runs on pattern recognition and a stubbornness about problems more than on being any kind of code guru.
You don't need to have built the model to be good at getting it to behave.
There's a gap between building a model and interpreting one, and that gap is where most AI work actually lives now. It's also the part companies keep mis-hiring for.
Look at the postings. An "AI strategy and enablement" role asks for five years of coding and a technical degree, then near the bottom, almost as an afterthought, one line about psychology, philosophy or design. The work is judgement and communication while the requirement is calibrated for something else.
A degree is cheap to verify. "Can reason about how a model behaves once it's loose in the world" is expensive to verify and easy to fake in a 45-minute interview. So companies filter on the signal they can read.
But what actually predicts the work is what someone has built, broken and fixed. Kojima and his co-authors took a large model and, without touching its weights or writing a line of code, added one instruction to the prompt: "let's think step by step." On a set of word problems the same model went from 17.7% correct to 78.7%.
That's the skill. Reading how the model behaves, knowing what to say to it, who it's talking to, and when it should say nothing at all.
The market hasn't caught up to this yet, and that's the part I find interesting. Interpretive fluency is hard to measure, so it's underpriced. Whoever learns to measure it from a portfolio instead of a transcript will hire better people for less. The mispricing is a real opportunity.
A technical credential and actual technical fluency aren't the same thing.
The people who close that gap have invested in the abstract, analytical side of this, and spent enough time watching these systems fail to know exactly where to look.
That's the work I've been doing, and it's the work I'd back myself on.
- Anka