Think Like Trixie
An eight-year-old asked what AI is, so we made her the hero of the answer.
How do you explain AI to a child?
Tell them a story where the AI is a friend who needs help, and let the child work out how it thinks. That’s what Think Like Trixie does. It’s a free picture book for ages 5 to 10 that I co-wrote. When her robot Trixie freezes mid-sentence, Vea shrinks herself down and dives into Trixie’s head to switch her back on, and along the way she learns what a prompt is, why some words count for more than others, and why the AI needs her to check its work. At the back, the book defines seven ideas in words a child can use, from tokens to human-in-the-loop.
Where the question came from
Vea is our cousin’s daughter in Singapore. Her question, “Mommy, what is AI?”, found its way to us, because building AI is what we do all day. The least efficient answer we could think of was a whole book.
What a child learns from it
Vea has an hour until her mom’s birthday party and no card written. Inside Trixie’s head she meets Modelle, a ball of golden light who says she’s a Large Language Model, and the Trixels, tiny pixel helpers who carry words around and build sentences. Each thing that goes wrong down there teaches one idea:
- Asked for “the perfect card”, the Trixels write three, and none of them sounds like Vea’s mom. A fuzzy prompt gets a fuzzy answer.
- Some word blocks are heavy and some are light. Heavy ones, like MOM, matter more to the sentence.
- The AI knows lots of words but not her family. Only Vea knows about the kaya-peanut butter disaster that made her mom laugh.
- The new card gets her mom wrong, and Vea catches it. A person stays on the team and checks what the AI makes.
- Modelle can think but can’t act. Trixie is the body that carries the card. Brain and body together make an AI agent.
The back of the book is Vea’s AI Discovery Dictionary, with seven entries: artificial intelligence, large language model, prompt, tokens, weights, human-in-the-loop and AI agent. A note for grown-ups suggests four things to talk about afterwards: ask clearly, check the answer, think before you share, and remember that AI isn’t a person.
Drawings that move as you read
Ophelia Chiu illustrated the book: the cover in full colour and every page in pencil. In the online book those pencil drawings move. Each one is cut into layers from her own lines, and the layers move in code as the page plays. A second copy of every line, nudged by less than two pixels, swaps with the first to make the slight shimmer of hand-drawn animation. The movement is made from her drawings, not generated by a model. The sound effects lettered into the drawings (CLANG, WHOOOOSH, BEEP! BOOP! BING!) play too, if you switch the sound on.
A read-aloud with AI voices
Choose “Listen to the story (beta)” on the title page and a narrator reads each page while its drawing moves in step. The word being read is underlined, and the page turns by itself at the end. Vea, Trixie and Modelle each speak in their own voice. Every voice is made with AI, and the book says so before the story starts.
We made the voices with two open models, run on our own machine. Kokoro reads the narration, and its own timing for each word is what moves the underline, so the mark lands where the voice says the word. Qwen3-TTS speaks the characters. Each character’s voice was designed once from a written description and kept as one reference recording, and every line copies that recording. When we designed Modelle’s voice afresh for each line, her pitch wandered by 12.8 semitones across eight lines. Copying one recording brought that down to 4.4.
The audio is recorded ahead of time and shipped with the book, page by page. The browser’s built-in speech was the easy option, but it sounds different on every device, can’t play the sound effects in step, and doesn’t report word timing reliably on an iPad. Recorded files sound the same everywhere and need no voice service.
Why it sits with my AI work
One of the book’s seven ideas is human-in-the-loop: Vea catches the card the AI got wrong. It’s the same idea behind the video annotation pipeline I led at Red Bull, where a person checks the moves the model isn’t sure of.
Read it
Book 1 is free to read on the LA TechSpace site, and it’ll stay that way. It opens in a browser with nothing to sign up for, and it remembers your page.