The Companion
Status: in development. Same core technology as the Talking Book; the plush form adds a safety-certification workstream that the book does not have.
A toy that never phones home. A plush that runs rhyming and counting games entirely on the chip inside it. The child presses the paw to talk. Their voice is processed where it is spoken and destroyed there. It is designed to work with the servers switched off, because there are none.
How it works
| Part | What it is | Why |
|---|---|---|
| The plush | Soft toy, hardware sealed inside | A form a toddler already holds, chews, and carries |
| The paw | Push-to-talk button | Mic is off until pressed. No wake word, no always-on |
| The chip | Microcontroller-class, under ten dollars | Runs the games on-device. No server, no subscription, no app. Measured numbers |
| The games | Rhyming, counting, category facts | Each maps to a named early-learning construct |
| Safety | Sealed battery, no small parts, materials certified | A toddler chews on it. Certification is a real workstream, not a footnote |
A plush that only listens when a child squeezes its paw cannot capture ambient audio. It also sleeps between presses, so the battery lasts months, not hours, and there is no charging cable to lose.
Scope is narrow on purpose. A general companion competes with cloud AI toys head-on. A plush that plays five rhyming games and counts to twenty competes with nothing, and its bounded model is faster, cheaper, and more coherent than a general one. Category-specific variants (an animal that narrates facts about itself) follow the same logic.
Why a toy that talks
- Back-and-forth is the mechanism. Romeo et al. (2018): in 36 four-to-six-year-olds, the number of conversational turns with adults, not the number of words heard, predicted Broca's area activation and verbal skill, independent of family income. 10.1177/0956797617742725
- The window is early. Gilkerson et al. (2018): conversational turns at 18 to 24 months accounted for 14% to 27% of the variance in IQ, verbal comprehension and vocabulary at ages 9 to 14. Adult word counts alone did not hold up once income was controlled. 10.1542/peds.2017-4276
- Guided play beats both free play and instruction. Zosh et al. (2018): when the child leads inside a structure an adult designed, learning goals are met better than by either extreme. A toy that answers the child's own rhyme or count is guided play by construction. 10.3389/fpsyg.2018.01124
- Children can learn language from a responsive machine. Kory-Westlund & Breazeal (2019), A Long-Term Study of Young Children's Rapport, Social Emulation, and Language Learning With a Peer-Like Robot Playmate. Frontiers in Robotics and AI. 10.3389/frobt.2019.00081. Young children built rapport and learned vocabulary from a peer-like robot. Those robots were cloud-backed and screen-faced, so this supports the direction, not our device.
The games sit on the same frameworks as the classroom app: the phonological-awareness ladder for rhymes, Gelman and Gallistel's counting principles for numbers.
"Companion" is a form factor, not a relationship claim
UNICEF's Policy guidance on AI for children (2021) sets nine requirements for child-centred AI including privacy by design, transparency, and never positioning AI as a substitute for human relationships. The classroom platform's rule that its AI never chats with a child comes from this document.
The Companion is a narrower case: a bounded game inside a plush. It does not remember the child, does not hold open conversation, does not act as a friend. It plays a rhyming game and counts apples. The word describes the shape of the toy.
Why no screen
- Hirsh-Pasek et al. (2015), four pillars: active, engaged, meaningful, socially interactive. A plush a child holds and talks to hits all four; a screen makes the first two harder for this age group. 10.1177/1529100615569721
- AAP (2016, revised 2025): shared use and developmentally healthy activity over solo screens; digital products designed around children's needs rather than engagement. 10.1542/peds.2016-2591 · 10.1542/peds.2025-075320
Why on the chip
Same two competitor failure modes as the book: Curio streams the child's voice to third-party AI operators by its own policy; Moxie went dead within days of its company closing. Ours has no such operator list because nothing leaves the toy, and no server to lose.
The only thing a parent has to understand is one sentence: it only listens when your child presses the paw.
The conversational-turn literature is about adults talking with children, and the social-robot literature used cloud-backed robots with faces. No published trial tests a bounded on-device plush with toddlers. We claim the mechanism and design to it; measuring the result is what the pilot is for.
The plush is the harder product: always in a toddler's hands and mouth, so materials, choking-hazard, and battery certification (CPSC, ASTM F963) are a real workstream and gate any sale. The book ships first for that reason.
Sources
- Romeo et al. 2018, Psychological Science. https://doi.org/10.1177/0956797617742725
- Gilkerson et al. 2018, Pediatrics. https://doi.org/10.1542/peds.2017-4276
- Harvard Center on the Developing Child, Serve and Return. https://developingchild.harvard.edu/key-concept/serve-and-return
- Zosh et al. 2018, Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2018.01124
- Kory-Westlund & Breazeal 2019, Frontiers in Robotics and AI. https://doi.org/10.3389/frobt.2019.00081
- Hirsh-Pasek et al. 2015, Psychological Science in the Public Interest. https://doi.org/10.1177/1529100615569721
- AAP 2016, Pediatrics. https://doi.org/10.1542/peds.2016-2591 · AAP 2025. https://doi.org/10.1542/peds.2025-075320
- UNICEF 2021, Policy guidance on AI for children. https://www.unicef.org/innocenti/reports/policy-guidance-ai-children
- Gelman & Gallistel 1978, The Child's Understanding of Number. https://www.hup.harvard.edu/books/9780674116375
- Curio privacy policy. https://heycurio.com/privacy
- Aftermath, Embodied/Moxie shutdown. https://aftermath.site/moxie-robot-ai-dying-llm-embodied/