What is “AI”?

So what IS "AI", when people say it right now? The first two primers took it apart: first the model, then the harness. Put them back together and you have what almost everyone means by "AI" right now: two things wearing one name. Here's the plain-English version of the pair, because once you can tell those two apart a LOT of this era stops being confusing.

The model is the famous part. The LLM, the large language model. It's a text predictor trained on more writing than any of us will ever read, then tuned on what people liked. It's brilliant...and it's frozen. It doesn't remember you, doesn't learn from yesterday, and wakes up brand new every single time. It predicts the next words. That's the whole job.

The harness is everything strapped around the model to make it useful, and it's the actual product you type into. It decides what gets loaded in front of the model each time (your files, the conversation so far, its standing instructions), what tools it can reach, what gets written down, and when the work counts as "done." ChatGPT is a model in a harness. Claude is a model in a harness. Every AI assistant or agent you use at work is the same shape: somebody's model, in somebody's harness.

Two more words you'll hear constantly, and they're simple once you have the split. An "agent" is a model whose harness lets it take steps: read a file, run something, check the result, go again. The looping is the harness. The deciding is the model. And "memory"...the model never remembers you. When it seems to, the harness is feeding it notes it saved from last time.

Here's why this matters for the question I keep asking, which comes next. The headlines are about models. Your day is mostly harness. When it "forgot" your instructions, the harness didn't load them, or a summary dropped them. When one assistant told a user who'd asked for help that "multiple critical flags have been submitted," that was a model talking way past its harness: it had no such ability, and the vendor confirmed that in writing. And the one from the harness primer that still surprises me: in most of the popular agent frameworks I checked, the default for "done" is that the model stopped asking for tools. That's it. The harness takes the model's word for it.

So here's the map as I see it. The model is rented, and it improves on the lab's schedule, not yours. The harness is what you configure, swap, and build on. And the record of your work, plus your judgment about what good looks like...that part is yours.

So when a lab promises a smarter model next quarter, notice which half they're talking about :)