The “AI experts” I meet fall into two camps: machine learning engineers who have never run a business, and business people who use AI constantly without understanding how it works. That gap will not resolve quickly, and waiting twenty years for today's ML graduates to acquire domain expertise is not a plan.
So: real competency, now. Not the latest social media post about the top ten ChatGPT tricks; what an LLM actually does, what persists between conversations and what doesn't, how tool use works, what RAG and embeddings are for, token economics, evaluation. The glossed-over nuts and bolts that make AI useful rather than magically impressive.
Karpathy describes LLMs as an emerging operating system — software 3.0 — which gave me language for something I had already been feeling. The Claude projects I build are not workflows, they are programs, and I would like to share them the way developers share code.
I am not closing the expert gap. I am documenting what I am actually doing, in enough detail that it might be useful to someone else. Including the parts that failed.
A specific problem, what I tried, what broke, and whether it was worth it.
A task that sounds trivial and isn't. Where the model's structural choices were better than mine, where they were confidently wrong, and how many passes it took before the output was something I would actually present.
I used my own learning-plan generator on a subject I knew nothing about, then followed the plan to see whether it held up. Write-up includes where the sequencing was wrong.
Longer pieces, only when I have something substantive to say.
A series of tests borrowing methods from psychology to probe how these models behave under pressure — with the caveat, stated in the piece, that behavioural resemblance is not evidence of an inner life.
What happens when a model is asked to notice that its interlocutor has stopped agreeing without saying so. The result surprised me enough to write it down carefully.
Downloadable. Take them, adapt them, no attribution needed.
The sequence I actually followed, not a reading list I aspire to.
What I am working on right now, including the things I have not figured out.