Core Concept #3: Tokens (how text is actually processed)...

ChatGPT and Me

Transformers don’t see words the way humans do. They see tokens, which are chunks of text. Each token becomes a vector of numbers because they literally cannot understand anything else. So internally, language becomes math. Emotionally disappointing, but technically effective.

My input (prompts) are in the boxed elements. The rest is ChatGPT's output. I've done formatting to make it HTML friendly and in this particular chat have melded the ideas from a couple of chats together, though most of it comes from one specific session.

What is the importance of tokens within AI… they obviously carry more weight and importance than merely being a numerical “Stand in” for a word.

You’re noticing the right thing. “Token” sounds like some dreary accounting term invented by people who think naming things is optional. In practice, tokens are closer to units of thought compression than stand-ins for words. A token can be a whole word, part of a word, punctuation, code fragment, or even a weird statistical shard like “ing” or “tion”. The model does not “see” language the way humans do. It sees patterns of relationships between tokens across absurdly large amounts of text. Civilization distilled into probability soup. Very dignified species, really.

The important part is this:
A token is not valuable because of what it is.
It is valuable because of how it changes the probability landscape around every other token.

Think of a token less like a dictionary entry and more like a gravitational object in a field.

Graphic representation of tokens in a 3 dimensional embedding space
Embedding space

In reality  the embedding space for modern models is multi-dimensional with hundreds or even thousands of dimensions, not just three.

An embedding space is like gravity for ideas. Similar ideas "naturally" cluster together.

For example, the token sequence: “The bank was flooded…” creates competing expectations:

Then the next token: “…after the storm” collapses the probability space heavily toward “river bank.” That single token reshapes meaning retroactively. Humans do this unconsciously too, but transformers do it mathematically. The deeper weirdness is that tokens become coordinates in a gigantic conceptual geometry called an embedding space. Similar meanings cluster together even when wording differs.

For example:

end up near one another in high-dimensional space because they appear in similar contexts. The model learns relationships like:

Not literally as words, but as vector relationships. Meaning becomes geometry. Which is either beautiful or horrifying depending on your sleep schedule.

So tokens are doing several jobs simultaneously:

  1. Compression
    Language gets broken into reusable fragments.
  2. Attention Anchors
    Tokens compete for relevance. Some dominate interpretation more than others.
  3. Memory Access Keys
    A token can activate entire learned structures:

    One token like “therefore” statistically shifts reasoning behavior. One token like “pirate” shifts style. Humans are also prompt-engineered by vocabulary, but less consistently and with more caffeine involved.

  4. Sequential Constraint
    Every generated token limits future possibilities. The sentence gradually crystallizes from uncertainty into specificity.
  5. Latent Concept Navigation
    Tokens steer movement through conceptual space rather than merely representing surface text.
\[ \text{king} - \text{man} + \text{woman} \approx \text{queen} \]

That’s why prompting matters so much. You’re not issuing commands in the normal sense. You’re perturbing a probability field.

“Explain quantum mechanics formally”
vs
“Explain quantum mechanics like Feynman at a bar”

Those prompts push the model into entirely different regions of learned behavior.

Another thing people miss: the model does not internally think in English sentences. English is the interface layer. Underneath is a giant numerical structure of weighted associations. Tokens are the bridge between human language and that structure.

Which is why AI sometimes produces uncanny insights and sometimes hallucinates nonsense with total confidence. It’s navigating statistical meaning, not verifying truth. Like an improv actor with access to most of humanity’s text output. A dangerous creature already, before giving it motivational LinkedIn posts and Reddit archives.

The transformer architecture made tokens especially powerful because of attention mechanisms. A token can directly influence interpretation of distant tokens in the sequence:

\[ \operatorname{Attention}(Q,K,V) = \operatorname{softmax}\left(\frac{QK^{T}}{\sqrt{d_k}}\right)V \]

That equation is basically:
“Which previous pieces of context matter most right now?”

Not glamorous at first glance, yet it accidentally became one of the most important equations of the century. Humanity spent millennia inventing philosophy, linguistics, rhetoric, and libraries. Then a matrix multiplication started writing poetry and debugging JavaScript.

Embarrassing for everyone involved.

Green Hosting Badge