Word Guessing Machines

Large Language Models (LLMs) like ChatGPT, Claude, and Gemini often sound like magical supercomputers with deep understanding and digital souls. But here’s a slightly less romantic take: at their core, these models are really just extremely good at guessing what word comes next in a sentence.

That’s it. No inner consciousness. No secret reasoning engine. Just a massive predictive text machine on steroids.

Let me explain.

When you ask an LLM something like “Write me a poem about a pirate who opens a coffee shop,” it doesn’t suddenly access a hidden pirate-entrepreneur database. It doesn’t know what a pirate is, or what a coffee shop is in the way we do. What it does is look at the prompt and try to predict—based on trillions of examples it has seen—what the next word probably is.

It’s like that autocomplete feature on your phone. But instead of giving you three suggestions and usually being wrong (I still think autocomplete is a horrible feature), this thing is trained on so much internet text that it can spin out page-long coherent outputs with surprisingly accurate structure and tone. All because it’s relentlessly, mathematically tuned to be good at guessing words.

Of course, that word-guessing ability is backed by some serious math. We’re talking about billions (or hundreds of billions) of parameters—essentially dials that get adjusted during training to make the model better at predicting those next words. It’s pattern recognition, not understanding. Like a parrot that’s memorized every TED talk and Reddit thread ever written.

You’ve probably heard that old thought experiment: put a million monkeys in front of a million typewriters, and eventually one of them will randomly bang out the complete works of Shakespeare. It’s absurd, but technically true—pure chaos eventually produces something genius… given enough time.

LLMs are like a much, much more efficient version of that. Instead of chaos, they’re trained on patterns. Instead of random banging, they’ve got algorithms trying to maximize likelihood based on past examples. But at the end of the day, they’re still just generating one word (or pixel) at a time, based on probabilities.

The difference? Monkeys get lucky once every few eternities. LLMs get lucky every few milliseconds.

What About the Image Generators?

Image generators like Midjourney, DALL·E, and Stable Diffusion work in a surprisingly similar way—but instead of predicting the next word, they predict the next pixel, or more accurately, the next chunk of visual noise that should evolve into something meaningful.

The newest wave of image models, especially diffusion models, start with pure static—literally just random noise—and then gradually “denoise” it over a series of steps. At each step, the model tries to guess what a less noisy version of that image might look like, conditioned on the prompt you gave it (like “a flamingo surfing a wave during sunset”). After 20 to 50 denoising passes, you’ve gone from fuzzy chaos to a fairly crisp, sometimes jaw-droppingly good image.

In other words: instead of guessing “what word should come next?”, these models are guessing “what should this blurry spot look like now?” over and over again until something beautiful (or weird, or terrifying) emerges.

And just like with LLMs, image models don’t understand what they’re making. They don’t know what a flamingo is, or what waves are. But they’ve seen enough training images with those labels that they can remix the patterns in a way that feels like art.

The results? Pretty wild. You can now generate fake stock photos, fantasy art, corporate logos, fake magazine covers, and illustrated children’s books with the same prompt-based magic as text. Though just like LLMs, the outputs can sometimes be uncanny, inconsistent, or accidentally hilarious—like when your “realistic man” ends up with twelve fingers and a half-melted face.

The generative process is still just a fancy form of pattern recognition and prediction. But it turns out when you scale up prediction to billions of parameters and train on the entire internet, you get something that feels like creativity.

Even if it’s really just math with good taste.

Prompt Used to create the image:

“A surreal, cinematic scene showing hundreds of monkeys in a vast, dimly lit room resembling a grand, old library or typewriting hall. Each monkey sits at a vintage typewriter, some pounding keys randomly, others looking confused or curious. Torn pages and half-typed manuscripts are scattered on the wooden floor. In the background, a faint glow highlights one perfectly typed page of Shakespeare’s “Hamlet” on a desk, as if by chance. The atmosphere is slightly whimsical and philosophical, with a soft golden light streaming in from high arched windows, casting dramatic shadows. 16:9 ratio, ultra-detailed, digital art, hyperrealistic style.”