Picture a Plinko board.
You know the one.
A chip drops from the top, bounces off a grid of pegs, and lands in a slot at the bottom that determines a prize.
It looks random.
It mostly is random.
Now imagine a version of that game that contains billions of pegs, and where the chip somehow learns from every single bounce before deciding where to fall next.
That is, roughly, how a large language model generates the next word when it’s talking to you.
What Is A Token?
Before the metaphor can land, you need one vocabulary word: token.
A token is not quite a word.
It is more like a chunk of language that the model has learned to recognize as a meaningful unit.
The word “running” might be one token.
The phrase “United States” might be stored as two.
A rare technical term might get broken into three or four fragments.
When an AI generates a response, it is not pulling out fully formed sentences.
It is predicting one token at a time, stacking them like Lego bricks until a complete thought has been built.
The Plinko Board Is the Neural Network
Here is where the metaphor starts earning its rent.
The Plinko board itself represents the neural network, a vast mathematical structure containing billions of individual parameters.
Think of each parameter as a peg on the board, and think of those pegs as being adjustable.
Some pegs push the falling chip left.
Some push it right.
During training, the model spent months reading enormous amounts of human text, and every time it made a mistake, the pegs were nudged slightly to prevent that mistake in the future.
By the end of training, the arrangement of pegs encodes something like the accumulated logic of a huge slice of human writing.
When you type a prompt, you are dropping a chip.
The chip represents your input, every word you wrote, every piece of context the model is carrying.
As it falls through the network, it hits peg after peg, and each interaction nudges it toward certain outcomes and away from others.
At the very bottom of the board, instead of dollar amounts, sit slots representing every possible token in the model’s vocabulary.
There might be fifty thousand of them.
The chip lands somewhere, and that landing is the model’s choice for the next word.
Then the process resets, the chip drops again, now carrying the new word along with everything that came before, and the next token gets chosen.
One bounce at a time.
Where the Metaphor Gets Interesting
Here is the thing about a real Plinko chip: once it’s dropped, physics takes over.
Every bounce is genuinely random within the constraints of gravity and geometry.
An LLM’s token drop is not like that at all, and understanding why is the part that changes how you see these systems.
The token is loaded.
Not randomly loaded, not arbitrarily loaded, but loaded in a way that reflects everything the model has learned about how language actually works.
When a sentence starts with “The capital of France is,” the token does not have an equal chance of landing on every slot in the LLM’s vocabulary.
The pegs have been arranged, through billions of training examples, to heavily favor “Paris.”
The probability is not certainty.
It is a dramatically weighted lean, the result of having seen that phrase and its answer too many times for the token to fall anywhere else.
But Plinko boards have unmoving pegs this is where the metaphor stops.
In a LLM, the pegs rearrange themselves during the fall.
As the model generates each new token, a mechanism inside the network called attention recalibrates the entire system based on what has been generated so far.
It is as if, between each token drop, a tiny crew runs out and physically adjusts the pegs to reflect the new state of the sentence.
If the model just generated “The dog chased the,” the pegs have already shifted to make “cat” more likely than “refrigerator.”
The whole board is responsive, not static.
Why This Matters More Than “It’s Just Autocomplete”
You have probably heard people dismiss AI writing as fancy autocomplete.
That framing is not wrong, exactly, but it is like calling a symphony “fancy noise.”
Yes, a language model is predicting the next token.
But it is doing so through a structure that has encoded enormous amounts of grammatical, factual, and contextual information into the arrangement of billions of parameters, and it is continuously updating that structure for every single token it generates.
The mechanism is genuinely novel.
It does not work the way your brain works.
It does not work the way a search engine works.
It is its own strange thing, and the Plinko board is just a doorway into seeing it clearly.
If you want to go deeper, two threads worth pulling: attention mechanisms, which explain how the board actually rearranges mid-game, and temperature, a setting that determines how “willing” the model is to land in a less likely slot.
Turn temperature up and the model gets creative.
Turn it down and it gets cautious.
Both are fascinating.
Neither is magic.
Just a chip, falling through a board that learned to care about where it landed.