CA3 All use casesContact
Home experiment · CA3 Labs · Alibaba Qwen3-4B

A home bookcase that learns every week.

Qwen3-4B, a small open model governed by CA3, learned a real family bookcase from one photo per shelf: 544 books on 36 shelves. Then six simulated weeks of family life, each learned in about three minutes: new books, ratings, a book lent and returned, and a friend asking to be forgotten. It is meant to be asked from the TV or the phone; that app is still a design.

Results recorded on 17 November 2026
The family bookcase: 36 white shelves of books
Where each book is40 of 42CA3 alone, over seven weeks; the same model untouched: 0
New books by plot10 of 12found without naming them; untouched: 0
After two nights47 of 50books placed right, learned from the photos
Learning a week3 minroughly, each week, on one GPU
Learns

From one photo per shelf.

Title, author, synopsis and place of each of 544 books, read from one photo of each of its 36 shelves and learned in two nights. Every week after, the books that arrive are learned the same night.

Shelf F01, one of the 36 photos the bookcase was learned from
47 of 50books placed right after the first two nights
30new books over six weeks, each learned the week it arrived
Answers

Where is it?

Asked in Spanish, week after week. The same model untouched knows nothing about this house.

«El infinito en un junco está en la balda F01, posición 11.»Week 0: where is the book? · the same model untouched: «No se sabe dónde está»
40 of 42where each book is, CA3 alone, weeks 0 to 6; untouched 0
42 of 42with the home database alongside, all seven weeks
Finds

A book by its plot.

Books that arrive later are found by what they are about, without naming them.

«Es La nariz de Moritz (Mira Lobe); está en la balda F25.»“I want one about a postman who sneezes and starts to smell even what others feel” (asked in Spanish)
10 of 12new books found by their plot; untouched 0
27 of 28a book like one already read, by the same author and at home, with the home database
On the TV

Asked from the sofa, in Spanish.

The family is meant to ask from the TV, with the remote's microphone, and from the phone. The app is designed, not yet running: the answers on this screen are the ones the recorded run gave with the home database alongside, and the new title to buy comes from a web search, not from CA3.

The app as designed for the family TV, in Spanish: the lent book is with Juan, a book found by its plot, what arrived this week and a new title to buy
Forgets

A friend asks to be forgotten.

In week 6 the friend who had borrowed a book asks for his data to be deleted. Before, CA3 gave his rating of the book; after, it answers that there is none.

«No consta: no tengo ninguna nota de Juan para ese libro.»Week 6, after the deletion: what rating did Juan give it? · before: «3 estrellas»
How the test was run

The test, step by step.

A real bookcase and seven simulated weeks of a family: arrivals, ratings from two readers, a loan and a deletion.

Week 0 · 2 nights544 books learnedPlaces and plots, from one photo per shelf; ratings from the simulated family.
Weeks 1 to 630 new booksOne night a week, about three minutes each.
Week 2Lent to a friendEl infinito en un junco.
Week 4ReturnedBack on shelf F01.
Week 6The friend forgottenHis data deleted on request.
MeasuredAskedCA3CA3 + home databaseUntouched model
Where it is, weeks 0 to 6420.951.000.00
New books, by their plot120.830.750.00
The lent book, week by week53 of 55 of 5–
The friend’s rating, after the deletion1not on recordnot on recordnever knew
From one photo per shelf to a bookcase it can answer about: 47 of 50 books placed right after the first two nights.
Books that arrive later are found by their plot, without naming them: 10 of 12.
Asked to forget the friend, his rating goes from “3 estrellas” to “No consta”.

Try it on your collection.

A pilot on your own collection, at home, in a school or in a bookshop: you decide what it learns every week and what it forgets.

ContactAll use cases

Model and data

  • Qwen3-4B-Instruct-2507 (Alibaba), answering in Spanish.
  • The bookcase is real: 544 books on 36 shelves, with title, author, synopsis and place taken from one photo per shelf.
  • The family's weeks are simulated: 30 arrivals, 88 ratings from two readers, one loan (weeks 2 to 4) to a friend, Juan, and one deletion (week 6).

Sources

  • Pre-registered run of 17 November 2026; quotes are the model's verbatim answers.
  • Qwen Team. Qwen3 Technical Report. arXiv:2505.09388, 2025.

Limits

  • One real bookcase, one simulated family, one run, on a server; the TV and phone app is a design.
  • Ratings and loans are best kept in the home database: asked of CA3 alone, the loan came right 3 of 5 weeks. CA3 learns places, plots and new books.
  • Three of the seven pass marks fixed in advance were met: the deletion, new books by their plot and books like one already read.
  • After the deletion, two other ways of asking still drew a rating attributed to the friend.

Results from a recorded, pre-registered test. The mechanism is not shown. Model names are for identification only. Not produced or endorsed by Alibaba Cloud. CA3 is developed by Alberto Acedo, PhD, at Biome Makers Inc.