Stop calling LLMs "AI".

A couple of years ago, while I was doing my MSc in applied data science at UMSI, we had the privilege of hosting Andrew Ng for an opening talk. I really look up to the guy - he founded and led Google Brain - and in one of his lectures he mentioned "AI is a marketing term for deep learning." That has stuck with me ever since.¹

...and then a few years later, he just keeps calling it AI.

My gripe with the term "AI" is I don't even know what artificial intelligence means. Intelligence, but fake? If we had to go there, fake intelligence honestly sounds better - at least it admits something about the product.

I'm pretty sure the brain doesn't work like a statistical model: layers of nodes, each one crunching through billions of ReLUs and passing the result to the next node, which crunches its own ReLUs, and so on, until somewhere in the soup a word comes out. (At least I think my brain doesn't work like that. Neuroscientists reading this: please let me know, I'd be very interested to be wrong.)

LLMs are fascinating - but all they do is predict the next token. Give them a sequence of tokens, they hand back a probability distribution over what could come next, we sample one, feed it back in, and repeat. That's it. The whole magic trick is a while loop.

What makes the thing impressive isn't the model - it's the software that handles the loop and the tool calls. What the young ones call a "harness" (and look, I mean, why do we keep inventing new terms? First we called it training, then it was fine-tuning - under the hood it's still calling .fit() on a loss, same as it ever was). The model is the engine. The harness, the data, the evals - that's the car.

This isn't a pedantry fight. A name should tell you what the thing is, so you know how to use it and where it breaks. "LLM" does that: it's a statistical model, it runs on tokens, it gets things wrong with total confidence, and it needs the same care you'd give any other component - a pipeline, retrieval, evaluation, guardrails. "AI" tells you nothing except that a marketing budget was spent on the product.

It matters most when the name is doing the selling. When a proposal says "AI" it's a vibe; when it says "an LLM answering over your data, with a retrieval pipeline and evals we run before anyone relies on it" - now you can actually evaluate it. We build these things for a living, and the clients who ask what is it actually doing are the ones we end up doing good work with.

I look forward to the day the world has moved on and LLMs fade away from the spotlight - the day "language model" is as boring as "database" is today. In the meantime, let's stop calling LLMs "AI".

*while downing a flat white, as is tradition