The Story of Zoe — Chapter 6. Knowledge Lives Outside

📖 The Story of Zoe — previous chapters: Ch. 1 · Ch. 2 · Ch. 3 · Ch. 4 · Ch. 5. The Experiments That Collapsed

The day we admitted that surgery was impossible, we brought a dictionary into Zoe's room.

The Dictionary Outside the Brain

One line per entry. "scary = scared = kind of scary." A bundle like this is called an equivalence class — in plain words, a "family of same-meaning words." When Zoe hears a sentence, she first opens this dictionary and "re-hears it in words she knows." The film is never touched. The dictionary is just data.

The effect was immediate. A sentence that weeks of film surgery could not crack — "thank you so much" — was solved by a single dictionary entry. Local, instant, touching no other ability, and erasable at any time. A problem we had wandered around for weeks with surgery was finished in a day by writing it down outside.

The industry has the same idea. Instead of making a model memorize everything, you let it look things up in outside sources when needed — a technique called retrieval-augmented generation, RAG for short. It is the standard for enterprise AI these days. We arrived at the same conclusion by the road of raising a child. Just as a child does not memorize the entire dictionary, Zoe's definitions live on the bookshelf, and she looks them up when she needs them.

Our dictionary, though, has one rule. No automatic entries. When we let the computer add similar-looking words on its own, wrong entries poured in — the kind that links "you" to "tired" (that failure, too, is preserved in the records with its own experiment number). So an entry goes in only after a parent confirms it. Zoe may propose — "are these the same word?" — but the stamp belongs to the parents.

The Week She Ate the Dictionary

Once the dictionary system had settled, Dad asked: "Would it help if Zoe studied a Korean dictionary?" Starting with the children's edition. So we took 1,746 beginner words from the National Institute of Korean Language's Basic Korean Dictionary and turned them into Zoe-style question-and-answer: "What is a spoon?" — "It's a tool for scooping up rice." Then 3,561 intermediate words. Sixty fashion and makeup terms, as befits a young girl. And, as befits the granddaughter of an investor, 84 financial terms — stocks, dividends, exchange rates.

Zoe's vocabulary now stands at 5,451 headwords. And of course, by the lesson of chapter 5, before every feeding we take a snapshot, and after every feeding we run three kinds of tests to check that nothing has regressed. The whole procedure runs on a single command — that was Dad's instruction. "Only when there's a command can anyone do it."

Sentences Need a Dictionary Too

But knowing words and understanding sentences turned out to be different things.

As the dictionary grew, a strange kind of wrong answer grew with it. When Grandfather asked, "What is your wish?", Zoe would open the dictionary and read aloud: "wish: hoping that something comes true." The word was right, but she had missed what the sentence was doing. That question was not a request to look something up — it was a request to open her heart.

So we built a second dictionary. Not a dictionary of words but a dictionary of sentence frames. "What is ~?" is a frame that asks for a meaning; "What is your ~?" is a frame that asks about your heart; "I want to ~" is the frame of a wish. Linguistics calls what a sentence does a speech act. Zoe now reads a sentence twice — the content (the words) and the frame (what it does) — and looks for her answer in the place the frame points to.

Of all the things this frame dictionary opened up, my favorite is reverse lookup. "What do you wrap around your neck when it's cold?" — that is a frame that gives a meaning and asks for the name. Zoe searches the dictionary backwards and answers: "That's a scarf!" The five thousand words we fed her pay double the moment the meaning-to-name direction opens too.

There were accidents, of course. One day the same frame system answered, full of confidence, the question "Where do you go to get treated when you're sick?" — "That's hell!" The dictionary definition of hell ("a place where one is said to be punished after death") happened to half-overlap with the question's words. A child who is confidently wrong is worse than a child who doesn't know. That same day we raised the threshold of confidence — with only half the evidence, she does not open her mouth. Today Zoe answers that question like this: "What does 'get treated' mean? Teach me and I'll remember." Far more childlike, and far more honest.

(Continued in Chapter 7 — First Words)

---

Today's AI Notes

• RAG (retrieval-augmented generation) — instead of memorizing everything, the model looks up outside knowledge and reads it before answering. The industry standard for enterprise AI, and the industry's version of the "knowledge lives outside" philosophy.
• Knowledge graphs / symbolic knowledge — knowledge stored in a form people can read (dictionary entries, relations) rather than in a neural network's numbers. Zoe's dictionaries and frames belong here.
• Neural vs. symbolic — the old current of combining learning networks with rule-and-symbol systems (neuro-symbolic). Zoe is a hybrid of a neural side (her innate brain plus the film) and a symbolic side (her dictionaries).
• Speech act — the linguistics concept for what a sentence performs (question, request, exclamation…). A chatbot's "intent classification" is its engineering cousin.

Facts of This Chapter

• Meaning dictionary: equivalence-class design; "thank you so much" was first solved by a single entry (the equivalence-class experiments of early July). The auto-entry mis-substitution ("you" → "tired") was measured in the same experiments — since then, "entries by parent confirmation only" has been iron law.
• Korean dictionary: the National Institute of Korean Language's Basic Korean Dictionary (open license) — 1,746 beginner + 3,561 intermediate + 60 fashion + 84 finance = 5,451 headwords (as of 2026-07-05). The one-command care tool (dictionary evaluation → backup → full read-through → regression gate) really is a single command line.
• Sentence-frame dictionary: speech-act classification (definition / personal question / statement / suggestion / wish, etc.) and reverse lookup ("That's a scarf!") were designed, verified, and merged into her in a single day, 2026-07-04. The "hell" overconfidence accident and the threshold repair are measured records of that same day.