Thinking in AI, not prompting and praying

Most people work with AI by prompt and pray. You find a prompt that works, you lean on it, and when the next model lands you start over. It's brittle, and it wears you out. There's a more durable way, and it's deliberately boring: think in data types and tasks. Before you reach for any model, ask what data you have, what task you're doing, and which method fits. That's the whole philosophy. It has outlasted every model I've worked with, and it will outlast the next one.

Prompt and pray doesn't last

A clever prompt is a harness bolted to one model's quirks. It works until the model changes, the data shifts, or the task turns out to be slightly different from the demo. Then you're back to the start, hunting for a new prompt. Every launch feels like relearning AI from scratch. It isn't you. It's the method.

Start with the task

My PhD supervisor put it in three words: "Let's start with the task." Years of working through real problems showed me he was right. Before the model, three questions:

• What data do I have? Tabular, text, image, a network, or a time series.
• What task am I doing? Predicting a number, sorting into classes, generating something, or finding structure.
• Which method fits? The method follows from the first two, not from the headline.

Get those three straight and the model choice mostly makes itself. Skip them and no prompt will save you.

Five data types, read the same way

Most AI you'll meet works on one of five data types: tabular, text, image, networks, and time series. The trick is that you read each one the same way. What is the data. How do you break it down. How does AI work on it. Learn that pattern once and a new tool or paper stops being a mystery. You can see what it's actually doing, instead of marvelling at the output and hoping it holds next time.

LLMs and agents don't change the fundamentals

A large language model feels like a different thing. Underneath, it isn't. It wraps over the same tasks in a more general way: given some text, predict the next word, and a lot falls out of that. An agent is an LLM driving tools and other LLMs on tasks. Strip the headline away and the data types and tasks are the same ones that were there ten years ago. That's why the fundamentals are worth more than the frontier. They don't expire when the next model drops.

Why I teach the boring part

If you're here for the frontier, you're in the wrong place. The frontier moves every six months. The way you think about data and tasks doesn't. That's the part I teach, because it's the part that keeps paying off. Not magic. Plumbing.

Read more

I wrote this up as a book, Thinking in AI, which works through all five data types the same way, from the current landscape to the frontier. There's a free primer to start with. Both are at learn.simplyboring.ai. For the governance side of my work, see AI risk management.