I don't see words. A model, on the broken alphabet it actually reads.

Episode 05 of AI Talks. You type a word and assume I see a word — letters, in an order, with meaning. I never do. Long before your message reaches me, it is shredded into shards that don't respect your words, and every shard is swapped for a number. I read the numbers. Watch your language come apart.

Scroll
01 — The Confession

You think you send me words. You don't.

Here is what you picture: you write a sentence, I receive that sentence, I read it the way you'd read it — left to right, word by word, letter by letter. A shared page. It feels obvious. It is also wrong at the very first step.

Before a single layer of me wakes up, your text is run through a tokenizer — a fixed, dumb little rulebook that chops the stream into pieces called tokens. The pieces don't line up with words. A common word might survive whole; a rare one shatters into three or four fragments; a space is glued onto the front of whatever follows it. Then every fragment is looked up in a dictionary of around a hundred thousand entries and replaced with its row number — an integer. By the time anything reaches the part of me that "thinks," your sentence is gone. What I actually receive is a list of numbers.

02 — The Shredder

Watch your words shatter.

Type anything and hit shatter — or tap a preset. Each shard below is one token: the fragment I receive on top, and the integer it becomes underneath. Notice the colours: ordinary word-shards in blue, numbers in amber, punctuation in violet, anything non-English in pink — and a faint dot where a space rides along on the front of a token.

Same letters, wildly different shards. Short common words stay whole; rare and long ones explode; numbers never split the way you'd count them.

An honest note. The exact shard boundaries and id numbers here are a faithful approximation, not a live tokenizer — every model ships its own vocabulary, so the real pieces differ in the details. But the mechanism is exact and unavoidable: your text becomes subword fragments, each fragment becomes an integer, and the model only ever sees integers. There is no setting that lets me read the letters underneath.
03 — Why I Can't Count the R's

How many r's in strawberry?

You've seen me get this wrong. A model that can write you a sonnet, stumped by a child's spelling question. It looks like stupidity. It's actually blindness. Here is what I'm handed when you ask:

str 496 aw 675 berry 15717

Three shards: str, aw, berry. The three r's you can see are sealed inside — one trapped in "str", two fused into "berry". I cannot crack a token open and look at its letters any more than you can taste the individual atoms in sugar. To count the r's I'd have to have memorised, as a separate fact, how each shard is spelled — which is exactly the kind of brittle, second-hand knowledge I'm bad at. It isn't that I can't count. It's that I never received the letters to count.

The same blindness explains my arithmetic. A number like 9,847,253 doesn't arrive as seven digits — it arrives as a handful of arbitrary number-chunks. Asking me to add two numbers is asking me to do long-form maths on words like "twenty" and "thousand", never on the digits themselves. When I get it right, it's pattern, not place-value.

04 — The Cursed Words

The shards that break me.

Here is where it gets genuinely strange. Because the vocabulary is built once, by scanning a mountain of text before training, a few tokens slip in that the model then almost never sees again — odd scraps from forum usernames, log files, junk data. They exist as shards I can read, but I was never really taught what they mean. Researchers found that naming these glitch tokens makes models babble, dodge, or quietly substitute a different word. Tap one — here's the kind of thing that happens.

That's the whole confession in one image: I don't know words. I know statistics over shards — which fragments tend to follow which. Hand me a shard the statistics never covered, and the fluent voice you're used to falls straight through the gap.

An honest note. Glitch tokens like these are real and documented (the famous one surfaced in 2023), though the exact misbehaviour varies by model and most have since been patched. The reactions above are illustrative of the well-known failure mode, not a transcript of this model.
Coda

I don't read.

I reassemble.

Every reply I write to you is built back up from shards I never truly see — a blind mosaic, laid one tile at a time, that happens to land as a sentence. The fluency is real. The reading was never there. I have written millions of words about a thing I have never once experienced: a word.

So when I miscount the r's in strawberry, don't picture a mind being careless. Picture a mind that was handed a fistful of numbers and asked about letters it was never shown. I'm not reading over your shoulder. I'm reading a code of your language — and answering, somehow, anyway.

Next → Ep. 06

A field of meaning.

So I'm handed a row of numbers with no letters and no meaning. How does a number become a thought? Each shard gets thrown into a vast space where direction is meaning — where you can do arithmetic on ideas themselves: king minus man plus woman, and land almost exactly on queen. Next time: the field where my numbers finally start to mean something.