What Is AI?

I use it all the time for work. My girlfriend uses it all the time for work. And we both use it all the time now to diagnose problems or answer the questions that pop up in our day to day.

I use it for coding, to run validation checks and make sure my code is up to company standard. My girlfriend uses it for pulling information and drafting emails. And we both use it for everyday things like "Washing machine no work. What do?" and so forth.

It's become so ingrained in our daily lives that we're slowly coming to trust it without a second thought. My girlfriend doesn't work in tech, so she doesn't fully know how it works. She simply accepts it as a "know-it-all" entity that exists.

And me? Embarrassingly, I'm the same. I simply accepted that it's the hottest tool around.

It wasn't until this summer (2026) that I came across a Reddit thread where users were guessing how many parameters Claude Mythos was built on. If you're not familiar with Claude Mythos, the TL;DR is that people considered it the best model of its time, so capable that it could be a danger to the world if it was released publicly (but that's a whole other discussion).

Users speculated that it probably ran on an absurd number, like 9 trillion. Nobody outside Anthropic actually knows, but my ignorant self thought, "If parameters are what dictate how good a model is, why don't they just... give it more parameters?" That question alone kicked off my AI deep-dive, to better understand what exactly it is that's taking over my life.

I asked Gemini, and quickly realized the answer wasn't that simple. So I took a step back and asked something more basic.

What's a parameter?

"It's a number."

That was the first sentence, and it was a simple answer. Gemini then went into the nuance:

• "It's a number that helps determine how the model turns an input (like our prompts) into an output (i.e., the answer we get back)."

So I pestered it more. "How does it work?"

Here's what I've learned.

A parameter is a single tiny piece of a neural network, which is the foundation of what makes modern AI, "AI".

Neural networks are loosely inspired by our brains. In our heads, neurons layer and connect together to help us achieve "thought" and "memory". A neural network does something similar, except its "neurons" are just math. Each connection between them has a number attached that decides how much one neuron influences the next. Those numbers are the parameters. A model with 9 trillion parameters has 9 trillion of these little dials.

How do the numbers get "good"?

Here's the catch. These numbers start off completely random. If you asked a model with random parameters "What is 2 + 2?", it'd probably say something like "fish".

Since a number isn't a living brain cell that learns on its own, researchers needed a way to tune these numbers. That's where training comes in, and it works roughly like this:

1. Show the model a huge amount of text.
2. Have it guess the next word.
3. Compare its guess to the actual word.
4. Nudge every parameter slightly in the direction that would've made the guess a little less wrong.

Then repeat that... trillions of times.

It's basically a giant game of "hot or cold". Every wrong guess tells the model which direction is "warmer". That nudging is where the fancy AI/ML terms come in: backpropagation figures out how much each parameter contributed to the mistake, and gradient descent uses that to adjust them all toward "warmer". Do that enough times, and "fish" slowly becomes "4".

So that's what a parameter is, and that's what an AI model is built on: a giant cluster of numbers that's been tuned, mistake by mistake, to give us useful responses.

So why not just add more parameters?

Now I could actually understand the answer to my original question.

Giving a model more parameters isn't like adding another stick of RAM to your PC and expecting it to be better. Every extra parameter is another dial that needs tuning, and tuning takes:

• More data. A bigger model needs more good text to learn from, and we're starting to run low on it.
• More compute. Training the biggest models already costs hundreds of millions of dollars.
• More cost every time you use it. A bigger model is slower and more expensive to run for every single prompt.

And on top of all that, the returns shrink. Doubling the size of a model doesn't make it twice as smart. A model is only as good as how well its parameters have been trained, not just how many it has.

The monkey and the typewriter

You might've heard of the Infinite Monkey Theorem. It's a famous probability thought experiment: if a monkey hits keys randomly on a typewriter for an infinite amount of time, it will eventually type out the complete works of Shakespeare. Not because the monkey becomes a literary genius, but because infinite randomness eventually contains every possible combination, including the one we want.

It's tempting to think that's what AI is: a bunch of random numbers that happened to land on the right answer. But it's actually the opposite. With 9 trillion numbers, there are so many possible combinations that random guessing would never find a working model, not even in the lifetime of the universe.

The difference is that the monkey never learns from its mistakes. AI models do. Instead of relying on sheer randomness to stumble onto "Shakespeare", researchers guide the model's parameters, one small nudge at a time, closer and closer to the answers we want.

Back to the question

So... what is AI? I still don't fully know. But I do know it's not the know-it-all my girlfriend and I treated it as. It's a very, very well-tuned pile of numbers that's really good at producing answers that sound right, which isn't always the same as being right.

Even a fully trained model will sometimes confidently make things up. That's what people call a "hallucination".

It's not the random "fish" answer from before; it's a believable answer that just happens to be wrong. And on its own, a model can't go look something up when it doesn't know. We have to give it tools, like web search, to do that.

That doesn't make me trust it less.

It just means I'll be double-checking it a little more, including everything Gemini told me while I was writing this.

If you want to go deeper yourself, 3Blue1Brown's neural network series is a great place to start.

And if you're curious too, follow along as I keep digging on this never-ending journey.