What does AI look like?

Think 2026

Turn agentic AI into real business value | Think keynotes

Artificial intelligence

IT automation

What does AI look like?

Long called “black boxes,” AI models are increasingly yielding their secret inner workings—even, this playful guide shows, in terms that anyone can understand.

Published 08 July 2026

By

David Zax

The stupid question

Recently, over coffee, a colleague relayed a question he’d been asked: what does AI look like?

I smirked. I muttered something about the tremendous complexity of “tensors” and “self-attention” and “neural nets,” words I had encountered from other, more technical IBM colleagues. Through my use of the jargon, I signaled the impossibility of meeting the question-asker on their Neanderthalic level.

More to the point, I thought, hadn’t everyone heard that AI was essentially a “black box”? I’d taken that to mean that we basically couldn’t know what AI looked like, under the hood. It was a fool’s errand.

We concluded our coffee, and I headed back to my desk. Obviously, if AI “looked like” anything, it was some impenetrable mix of complexity and opacity. “‘What does AI look like?’” I thought, smugly. “What a stupid question!”

That night, I stared at the ceiling sleeplessly, my mind racing. I wanted to know what AI looked like.

Sure, it was a “black box.” But humans had made it, so surely we knew something about what it looked like?

Disconnected images flitted through my brain. I knew in a vague way about data centers far away. I’d seen schematics of “neural nets,” complete with their “layers.” But what were the layers made of? And where were they? In the data centers? Or inside my computer, in the support files of my ChatGPT application? And was AI written in code? It had to be—all computer stuff was code!

I rose from my bed and commenced to do what one does in 2026. I spent an hour pestering ChatGPT about what it looked like, asking it to walk me through the major research papers that had introduced it to the world, then begging it to “explain like I’m five.” Finally, I reached a tentative conclusion: sure, if you scratched away at normal old-fashioned software, you got code; but if you scratched away at AI, what you got was … numbers. Just lots and lots of … numbers. These numbers were arrayed in massive grids: neat rows and columns.

So AI was basically … spreadsheets? “Yes, that’s fair,” said my robot interlocutor. “Spreadsheets with special plumbing.”

I closed my computer. Mystery solved: AI looked like spreadsheets. For the rest of the night, I slept soundly.

The next day, unease crept back in. OK, sure. AI looked like huge spreadsheets full of numbers. But upon reflection, I had just a tiny follow-up question ...

WHAT???

I had known spreadsheets, even fairly large ones. But none of them had been able to talk to me. So what was going on with these spreadsheets that made them so smart?

Now that I had some idea what AI looked like, I wanted to understand what it looked like in action—how, under the hood, it all worked.

Was I deluded in thinking that might be possible? I’d recently read a pronouncement from Anthropic CEO Dario Amodei. He’d written: “People outside the field are often surprised and alarmed to learn that we do not understand how our own AI creations work.”

Even so, I thought, understanding is relative, and I had a feeling Amodei knew more than me. I was willing to bet he even knew the thing about the spreadsheets.

So AI looked like giant spreadsheets full of numbers. But I had a tiny follow-up question... WHAT???

So, over the next few weeks, I embarked upon a challenge: could I—a writer who had never written a line of code, who had taken the bare minimum of math in college—approach a practitioner’s understanding not just of what AI looked like, but why?

To my surprise, I discovered that despite the “black box” label, we actually can see inside a large language model (LLM). We know what AI looks like—and with each passing day, researchers are learning more about just how it works. By following our curiosity, even laypeople like you and me can begin to bring the picture into focus.

The more I learned, the more I came to believe not only that we can see and understand AI, but that we must. In fact, I became convinced that the ongoing effort to understand AI—by both experts and laypeople alike—will be the very thing that keeps it safe. Let’s begin.

The latest AI trends, brought to you by experts

Get curated insights on the most important—and intriguing—AI news. Subscribe to our weekly Think newsletter. See the IBM Privacy Statement.

Thank you! You are subscribed.

The big picture

A modern AI system like ChatGPT (or Claude, DeepSeek or Granite) is a series of spreadsheets filled with numbers. It’s a curious irony that this verbal technology is, at root, almost purely numerical.

How many spreadsheets? It depends on the model. For the rest of this article, we’ll focus our energy on OpenAI’s GPT-3,...