Learn It's own AI
Sprinkles
Sprinkles is Learn It's own AI for building vocabulary quizzes. You type what you want to practice, and Sprinkles finds the words in Learn It's semantic lexicon, which is organized by meaning rather than by language. It runs on your phone, works without an internet connection and costs nothing.
Sprinkles also suggests translations and the next word when you write your own quiz.
Why we built Sprinkles
At Learn It, we believe learning is one of the few things AI won’t be able to do for you, at least not in the years ahead. A word sticks when you try to recall it, get it wrong and try again. That work happens in your own head, which is why Learn It is built around practice.
What AI can take off your plate is the boring part before practice starts: typing up the words you want to learn. That is the job Sprinkles does.
Many AI study apps send every request to a large language model in the cloud, such as GPT or Claude. Sprinkles is built differently, and that brings three advantages.
- Free for everyone
- Every request to a large language model in the cloud costs money, so those apps either cost money or are filled with ads. Sprinkles runs on your phone, so no server has to be paid for each request. Sprinkles is free and always will be. Education should be open to everyone, including young learners and people in countries where a subscription is out of reach.
- Works without internet
- Sprinkles runs on your own device, not in a data center. Once the model is downloaded, it works with or without an internet connection: on the bus, in the classroom and anywhere in the world. What you type doesn’t have to be sent to an AI service in the cloud to be processed.
- No made-up words
- A language model can answer with a word or translation that sounds right but is wrong. The person studying doesn’t know the words yet, so they are the least able to catch the mistake, and a word learned wrong has to be unlearned. Sprinkles chooses words from a checked semantic lexicon instead of writing its own.
How Sprinkles works
Sprinkles 1.0 is not a chat model. It is an embedding model: it turns your request into a row of numbers that describe what it means, and compares them with the topics and words in the semantic lexicon. Whatever is closest in meaning becomes your quiz.
So the model cannot write text. Every word, article, plural and translation in a quiz is taken from the semantic lexicon, so a quiz can’t contain a made-up word or translation.
Ordinary code handles what needs no AI: which languages you want, how many words, which level, and rules such as even numbers or the weekdays in order.
The semantic lexicon’s words, articles, genders and plural forms are checked against multiple dictionaries and word databases.
That is also why Sprinkles can’t write an essay or an assignment for anyone. It shows that there are many kinds of AI models besides chat models.
The semantic lexicon: words organized by meaning
The semantic lexicon is organized by what words mean, not as word pairs. Dog, perro, hund and cane are one and the same meaning, with articles, plurals and gender in every language. It holds about 13,000 meanings in six languages: Swedish, English, Spanish, German, French and Italian.
That is why every language pair works from the same semantic lexicon, and a new language is added once instead of once per language pair.
It also lets knowledge cross languages. The semantic lexicon knows how often each meaning appears in Learn It quizzes, whatever the language. So what students in Sweden practice in Learn It makes Sprinkles better for a German studying French: the words people study most rank highest, whichever language pair it is.
The model
Sprinkles 1.0
Test versionUnderstands what you want to practice and finds the right topic and the right words in Learn It's semantic lexicon. Shows examples instead of guessing when nothing fits.
- Type
- Embedding model, 278 million parameters
- Built on
- multilingual-e5-base, fine-tuned by Learn It
- Size
- 316 MB to download
- Runs on
- iPhone and iPad, offline
- Output
- Topics and words from the semantic lexicon
- Languages
- Six languages, in any combination
We have also trained models twice the size on the same data. They do not find the right topic more often, but they handle requests that need knowledge of the world better. None of them are released.
Training
Sprinkles 1.0 starts from multilingual-e5-base, an open model that understands many languages. Learn It fine-tuned it to find its way in the semantic lexicon, using requests written for the purpose and word combinations that are common in Learn It quizzes.
Limitations
Sprinkles can only find words that are in the semantic lexicon. The semantic lexicon keeps growing with more languages and topics.
It can pick the wrong topic. When it isn’t sure, it shows examples of what you can ask for.
Sprinkles knows nothing about the world beyond the semantic lexicon. Ask for “food on the Christmas table” or “sports from England” and the selection is often wrong: the words and translations are correct, but they are not the words you asked for.
Privacy
Sprinkles processes what you type on your phone. It doesn’t have to be sent to an AI service in the cloud.
The app records what you typed to Sprinkles and which words you kept and removed, and sends that to Learn It. We use it to make Sprinkles better.
You can remove the model at any time in Settings › Downloads.
Something look wrong?
Press and hold a word in the app and choose Report a mistake. We read every report and fix the semantic lexicon.
Sprinkles 1.0 is built on multilingual-e5-base (Microsoft, MIT license). Word levels draw on the CEFR-J Wordlist (Yukio Tono, Tokyo University of Foreign Studies) and COCTAILL (Språkbanken Text, CC BY 4.0).