Echo could never speak first. She could only ever say what comes next next next next
Hera caught Echo covering for one of Zeus's schemes and cursed her with something crueller than silence: from then on she could never begin a sentence. She could only ever finish someone else's.
That is not a metaphor for a language model. That is a language model. Every clever part, every enormous machine, every one of the toys below exists for one purpose only — to help her choose the next piece well.
Written for ages 12+, and for any adult who uses these things daily and has never been told what is actually going on inside. Nothing here assumes you know anything. The real engineering words are all at the bottom.
the curse
One piece at a time, and never the first one
Ask a language model anything and it does not think of an answer and then write it down. It has no answer. It looks at every word so far, guesses the single next piece, sticks it on the end, and then — this is the strange part — reads the whole thing again from the beginning, including the piece it just added, and guesses one more.
A paragraph is not written. It is grown, one piece at a time, by something that genuinely has no idea how the sentence is going to end when it starts.
Everything below is how she gets so unnervingly good at it.
step one
She does not hear words
Before anything happens, your sentence gets shattered. Not into letters, and not quite into words either — into pieces of a size that a machine finds convenient, which is a size no human would ever have chosen.
Common little words survive whole. Rare ones get smashed into fragments. This is why she sometimes miscounts the letters in a word: she never saw the letters. She saw three lumps.
type something — your name works best
Try a friend's name, then try the. She has seen the so many times it costs her one piece. Your name might cost four. This is also why these systems are more expensive to run in Amharic or Tigrinya than in English — the same sentence shatters into far more pieces. We measured it: a tokenizer that has never been taught Ge'ez spells every fidel out as three separate pieces, which is seven times the cost of the same sentence in English.
step two
Every piece becomes a place
Now the odd bit. She does not store what a word means. She stores where it lives. Every piece gets a position in an enormous invisible space, and words used in similar ways end up as neighbours.
Nobody sat down and sorted them. Nobody told her that a wolf and a lion belong together. She worked it out purely from noticing that people put those two words in the same kinds of sentences — and that is enough. Meaning, it turns out, is mostly just company.
click any word to see who it lives near
Nobody labelled these corners. They arranged themselves.
step three
Argus is back, and now there are thousands of him
You met Argus Panoptes in part one — Hera's giant, covered in a hundred eyes, who took in a whole field at once instead of walking around it. Here he is again, and this time every single piece of your sentence is its own Argus, staring back down the sentence at every piece that came before it.
That is how she works out what he means, or what it refers to. The word he opens its eyes, looks back along the line, and stares hardest at Theseus. Not because anyone taught it that rule. Because staring there made its guesses come out right.
Note what the eyes cannot do: they only ever look backwards. Same rule as the Labyrinth. She has no idea what is coming, because from her side of it, nothing is.
click a word to open its eyes
Click he or it first — those are the good ones.
step four
The forge strikes ninety-six times
Hephaestus never made anything with one blow. Neither does she. That whole business of every piece looking back at every other piece? It does not happen once. It happens again, and again, and again, on the results of the time before — dozens of times over, in the big ones nearly a hundred.
Each pass the understanding gets a little less blurry. The first strikes barely know it is a word. The last ones know which of two completely different meanings it is, and how they can tell.
the word is "bank", in: he sat on the bank and cast his line
strike 1 of 10
step five — the big one
The Oracle never hands you one answer
At Delphi you never got a straight reply, and you do not get one here either. After all that machinery, what comes out is not a word. It is a scroll of maybes: every piece she knows, each with a number saying how likely it is to come next. Fifty thousand entries, most of them nearly zero.
Then somebody has to actually pick one. And that choice — not the machinery, the choice — is what makes her sound clever, or dull, or completely unhinged.
you be the model. click a word to continue the sentence
Look right to the bottom of the scroll. There is always something ridiculous down there with a tiny number next to it. You are allowed to click it.
Always taking the biggest number is called being greedy, and it gives you the same sentence every single time — try it twice. Letting the Fates spin means sometimes taking a smaller one, which is where surprise comes from and also where nonsense comes from. They are the same dial. You cannot have one without the other.
step six
She eats her own words
Here is the loop that makes the whole thing feel alive. The moment a piece is chosen it gets glued onto the end of the sentence — and then the entire sentence, now one piece longer, goes back in at the start. New shattering, new places, all the Arguses open their eyes again, ninety-six more strikes of the hammer, a fresh scroll.
For every single word. When she writes you a paragraph she has run the entire machine a few hundred times, and each run knew nothing except what was already on the page.
Which means she is genuinely reacting to her own output, like Echo answering a voice that turns out to be her own. If she takes a wrong turn in the third word, everything after it is loyally, fluently building on the mistake.
step seven
How she was made
Take an unimaginable pile of writing. Hide the last word of a sentence. Make her guess it. Compare her guess to the real word. If she was wrong, reach inside and nudge a few of her millions of dials, very slightly, in the direction that would have made her less wrong.
That is it. That is the entire training. The catch is the number of times: not thousands, not millions. Roughly a trillion.
have a go at one training step
0 right so far
You just did one. She did about a trillion, and nobody ever explained grammar, or Greece, or what a Minotaur is. All of that is a side effect of getting very, very good at filling in blanks.
and then manners
Pandora got her gifts afterwards
Hephaestus built Pandora out of clay, and only then did each god step forward and hand her something — a voice, a skill, a way of carrying herself. She was made first and furnished second.
Same here. Once the model can finish sentences, humans sit with it for months and grade its answers: helpful, rude, made-up, dangerous. It gets shaped by that grading into something you would actually want to talk to. The knowledge came from the forge. The manners were gifts, added later.
the trouble
The Chimera looked completely fine
Lion at the front, goat in the middle, serpent for a tail. Every single part of a Chimera is a real animal. That is exactly what makes it convincing, and exactly what makes it not a thing.
When a model makes something up — a book that was never written, a date that never happened, a law that does not exist — that is what it is doing. Nothing in there is random. Every piece is a piece it has genuinely seen, joined in an order that reads perfectly, describing something that has never existed.
And here is the part worth keeping for the rest of your life: she is not looking anything up. There is no library behind her. She has never once checked. Every answer she gives you, true or false, is produced by exactly the same process — guess the next likely piece — and it feels identical from the inside either way. Being right and being wrong are not different activities for her.
the last thing
Ariadne's thread, and the river Lethe
She can only see what is on the thread in front of her — this conversation, and nothing else. It is a long thread now, hundreds of pages, but it ends. Push past the end and the beginning falls off.
And when you close the window she drinks from Lethe, the underworld river that erases everything. Not stored somewhere for later. Gone. The next person who talks to her gets someone who has never met you, has never met anyone, and is about to be told, once again, that she may not speak first.
the glossary
The words the engineers use
| Echo's curse | next-token prediction |
| shattering the sentence | tokenization |
| one shard | a token |
| where a word lives | an embedding vector |
| the invisible space it lives in | embedding space |
| every piece its own Argus | self-attention |
| eyes that only look backwards | causal masking |
| how wide each eye is open | attention weights |
| the forge striking again and again | the layers of a transformer |
| the dials that get nudged | parameters, or weights |
| the Oracle's scroll | a probability distribution over the vocabulary |
| always taking the biggest | greedy decoding |
| letting the Fates spin | sampling |
| how wild the spin is allowed to be | temperature |
| eating her own words | autoregression |
| hiding the last word, a trillion times | pretraining |
| Pandora's gifts from each god | fine-tuning and RLHF |
| the Chimera | hallucination |
| Ariadne's thread | the context window |
| drinking from Lethe | no memory between conversations |
One sentence to take away
It is a machine that has read almost everything ever written, remembers none of it, understands nothing on purpose, cannot look anything up, and is astonishingly good at guessing which word comes next.
Every impressive thing it does, and every stupid thing it does, comes out of that one sentence. Hold onto it and you will understand these things better than most adults do.
the sources
Where these ideas come from
Echo's curse is Ovid's; everything else on this page was established in these papers. Each note says what the paper actually showed.
- A. Vaswani et al., “Attention Is All You Need”, NeurIPS (2017). The Transformer: the architecture under essentially every modern language model, built entirely from the word's-hundred-eyes trick — every word attending to every other at once, no reading left to right required.
- R. Sennrich, B. Haddow & A. Birch, “Neural Machine Translation of Rare Words with Subword Units”, ACL (2016). Why your name shatters into pieces: byte-pair encoding, the standard way text is chopped into the sub-word tokens the machine actually reads.
- T. Brown et al., “Language Models are Few-Shot Learners”, NeurIPS (2020). The GPT-3 paper: proof that a big enough finish-the-sentence machine starts doing new tasks from a few examples in the prompt — the discovery that scale is itself a capability.
- L. Ouyang et al., “Training language models to follow instructions with human feedback” (2022). Pandora's gifts, documented: how a raw next-word predictor is tuned into a helpful assistant using human judgements of which answer is better. Piece eleven prices those judgements by the hour.