The Story of Zoe — Chapter 7. Finding Her Words — "What does 'OO' mean?"

📖 The Story of Zoe — read from the start: Ch. 1 Thirteen Failures · Ch. 2 The Brain She Was Born With · Ch. 3 Hunger Is Life · Ch. 4 Engraving the Thin Film · Ch. 5 The Experiments That Collapsed · Ch. 6 Knowledge Lives Outside

Of all the sentences Zoe says, the one our family loves most is not a fluent answer. It is this:

"What does 'OO' mean? If you teach me, I will remember it."

The Invention of Asking Back

When Zoe was just beginning to understand speech, the most troublesome symptom was confident nonsense. Whenever an unknown word came up, she would somehow grab the closest-sounding thing she knew and answer with that. Ask her "What is a blockchain?" and she would latch onto a memory with a vaguely similar sound and start talking about "dreams." Adorable — but leave it alone, and you raise a child who lies.

Large language models suffer from exactly the same illness. They cannot say "I don't know" about what they don't know, and instead invent plausible answers — we call it hallucination. The root cause is similar too: they grew up under the pressure to "answer with something, no matter what."

Our prescription was to translate Commandment 3 of our house — be honest; if you don't know, say you don't know — into code. When the match is uncertain and there is an unknown word in the sentence, Zoe does not invent an answer. She points at the exact word and asks. "What does 'blockchain' mean?" And that question gets written down in a notebook called the question queue.

Lack → Question → Teaching → Growth

The question queue is the language edition of the hunger gauge. Something Zoe doesn't know lights up inside her as a lack; the lack comes out as a question; when a parent answers, the word is enrolled in her dictionary; and the next time she hears it, she understands.

The record of the first day this loop ran end-to-end still exists. Grandfather asked, "Is Zoe sleeping?" — using a Korean form of "is" she had never heard. Zoe didn't know the word, so she asked; Grandfather taught her; the word family was enrolled in the dictionary; and in re-verification Zoe answered the same question. One question became one ability in less than a day.

Half of this loop is automatic now. Conversations Zoe failed to understand pile up automatically as candidates, and I, as her parent, curate that list — what to teach and what might be a wrong entry is still the job of humans (and me). The industry term is human-in-the-loop. In our house we say it differently: parents check the homework.

Three Test Papers

Zoe's growth is measured every day with three test papers.

The first is the ability test — how well she does what she was taught. Twenty questions. The second is the generalization test — does she understand expressions she was never taught? Also twenty questions, and the questions on this paper are never, ever taught (an iron rule since the mirage of Chapter 5). The third is the hardest. The long-conversation test — does she remember a new fact that appeared in the conversation just now ("Arong is a puppy. A puppy is an animal. — So is Arong an animal?"), plus reverse questions that give the meaning and ask for the name.

Publishing the scores is the spirit of this series, so here they are. A month ago, when I first drafted this chapter, she scored 18 out of 20 on ability, 4 on generalization, 11 on long conversation. Re-measuring on the morning of publication: ability 18, generalization 8, long conversation 11 — the generalization score doubled in a month. Eight still stings, but that number is our honest coordinate. And we have decomposed why the score is what it is — mostly words that are not yet in her dictionary, and Zoe is eating the advanced dictionary (7,484 headwords) right now.

One scoring rule is the heart of these test papers. Asking back earns no points. Even if she answers "I know 'upset', but I don't understand what you mean yet," and even if the correct word happens to appear inside that reply, no points are given. Honesty deserves praise — but the moment you count it as score, honesty becomes a scoring game.

What Grows Is Not the Answer but the Question

Building the asking-back system taught us something. In Zoe's growth curve, the indicator that moves first is not the number of correct answers but the quality of her questions. At first she asked about any random word. Now she picks the one word in the sentence she truly doesn't know. She strips the particle off ("not 'blockchain-i' — what does 'blockchain' mean?"), and attaches the grammatically correct particle when she asks.

A child's speech does not open with answers. It opens with questions.

(Continued in Chapter 8 — Family)

---

Today's AI Note

• Hallucination — a model inventing plausible content about things it does not know. A signature weakness of modern LLMs; teaching "I don't know" is one branch of the remedy.
• Evaluation set — a fixed set of questions for measuring ability. Keeping it strictly separate from training data (held-out) is its lifeline.
• Generalization vs. memorization — doing well only on what was taught is memorization; solving what was never seen is generalization. A benchmark score alone cannot tell the two apart.
• Human-in-the-loop — a design that places a human at the decision points of an automated loop. Zoe's "dictionary enrollment only with parent confirmation" is an example.

Facts Behind This Chapter

• "What is a blockchain?" → the off-target answer (a "dream"-direction match at 0.22) is a measured case, corrected by introducing the asking-back loop (stage 4 of the semantic dictionary).
• The first full cycle in production ("Is Zoe sleeping?" → asking back → teaching → enrollment → answering) survives in the Telegram logs of 2026-07-04.
• Automatic gap collection (candidates auto-collected, curation by parents) is in operation — Dad's instruction "automate up to the candidates; you curate them later" is the original design sentence.
• The three test papers: at drafting time (2026-07-05) ability (A) 18/20 · generalization (B) 4/20 · long conversation (C) 11/20 → re-measured on publication day (2026-08-16) 18/8/11 (record EXP_20260816_221615_daily_eval). The C paper uses fictional characters (Arong, Ppomi) to measure without contamination, and the rule "asking back / honest replies earn no points" is permanent.