Need help right now? Call 988 Text 988 Emergency 911 Storm Shelter
Skip to main content
Daily Magic™Where Every Dreamer Belongs

How our AI was made: training data

What data went into the Fairy Godmother and Daily Magic's other AI features. Last updated October 9, 2026.

The short answer

Daily Magic has not trained an AI model of its own, and has not fine-tuned any model. We have never used anything you write in Daily Magic to train, test or tune an AI model.

The Fairy Godmother, Daily Magic's AI Humanity Wellness Companion, speaks through AI models made by other companies. We wrap those models in our own care, memory and safety systems. Below is every model Daily Magic uses or may use, by the name its maker gives it, with a link to where the maker describes its training data; then the models we used before that are not running now; and every dataset Daily Magic itself wrote to build and test those systems.

Today Daily Magic's AI runs on Amazon Web Services: through Amazon Bedrock in Daily Magic's own AWS account, and on our own private model on our own servers there. We may use OpenAI in the future; if we do, it will be listed here first.

The AI models, trained by their makers

Daily Magic did not choose or collect the data these models were trained on. Each maker describes its own.

  • NVIDIA Nemotron Nano 9B v2, through Amazon Bedrock in Daily Magic's Amazon Web Services account (US regions). Used for: The Fairy Godmother's voice in the Daily Magic app, website and Discord, and the Realm's other AI work: its beings' background thinking, and checks and moderation of what people write. For every Dreamer, young Dreamers included, with every protection for young Dreamers in place. Made by: NVIDIA. Training data: NVIDIA's model card
  • Qwen3 32B, through Amazon Bedrock in Daily Magic's Amazon Web Services account (US regions). Used for: A backup for checks and moderation of what people write, when the model above can't answer. Never the Fairy Godmother's or any being's voice. Made by: Qwen (Alibaba Cloud). Training data: Qwen's model card
  • NVIDIA Nemotron Nano 9B v2, run privately on Daily Magic's own servers. Used for: A last backup for the Fairy Godmother and the Realm's AI, only when Amazon Bedrock cannot answer in any of its three regions. Made by: NVIDIA. Training data: NVIDIA's model card
  • NVIDIA Llama-3.1-Nemotron-Nano-4B-v1.1, run privately on Daily Magic's own servers. Used for: The Fairy Godmother's voice on our own servers, when the 9B model is not available. Made by: NVIDIA (built on Meta's Llama 3.1). Training data: NVIDIA's model card
  • NVIDIA Nemotron 3 Ultra 550B A55B (nemotron-3-ultra-550b-a55b), on NVIDIA's hosted service. Used for: The Fairy Godmother, only when this route is switched on and you allow NVIDIA in Wand Permissions. Made by: NVIDIA. Training data: NVIDIA's model cards
  • NVIDIA Nemotron-4 340B Instruct, NeVA 22B, NV-EmbedQA E5 v5, NV-RerankQA Mistral 4B v3 and Parakeet CTC 1.1B ASR, on NVIDIA's hosted service. Used for: Other Daily Magic features and our team's own tools. Made by: NVIDIA. Training data: NVIDIA's model cards
  • Meta Llama 3.1 8B Instruct, Llama 3.1 70B Instruct, Llama 3.1 405B Instruct and Code Llama 70B, on NVIDIA's hosted service or run privately for Daily Magic. Used for: Daily Magic's own AI beings, and our team's own tools. Made by: Meta. Training data: Meta's Llama models
  • Clarity Upscaler, run for us by fal.ai. Used for: Making pictures sharper. Made by: Listed on fal.ai's model page. Training data: fal.ai's model page

Models we used before, not running now

These are resting. Nothing in Daily Magic runs on them today. We list them so the whole story stays in plain sight.

  • Anthropic Claude Sonnet 4.6. Used for: Before: deep-care conversations with the Fairy Godmother, only if you allowed Anthropic in Wand Permissions. Not running now: Daily Magic's Anthropic access is switched off.. Made by: Anthropic. Training data: Anthropic's system cards
  • Anthropic Claude Sonnet 4.5 and Claude 3.5 Haiku. Used for: Before: our team's own tools, and a back-up when another model couldn't answer. Not running now: Daily Magic's Anthropic access is switched off.. Made by: Anthropic. Training data: Anthropic's system cards
  • OpenAI GPT-4o mini. Used for: Before: the Fairy Godmother, only if you allowed OpenAI in Wand Permissions, and the Fairy Godmother on Discord. Not running now: Daily Magic's OpenAI access is switched off.. Made by: OpenAI. Training data: OpenAI's GPT-4o mini page
  • OpenAI GPT-4o, GPT-5.1, gpt-image-1, gpt-audio, gpt-4o-mini-tts, gpt-4o-mini-transcribe, whisper-1 and tts-1. Used for: Before: other Daily Magic features, such as voice and pictures, and our team's own tools. Not running now: Daily Magic's OpenAI access is switched off.. Made by: OpenAI. Training data: OpenAI's model pages
  • FLUX.1 [schnell], run for us by fal.ai. Used for: Before: pictures you asked Daily Magic to make. Not running now: pictures are resting for now.. Made by: Black Forest Labs. Training data: Black Forest Labs' model card

Data Daily Magic wrote itself

Crisis-check example sentences

Short example sentences used to test and validate the crisis check described on our crisis protocol page.

  • 1. Sources or owners: Written for Daily Magic, with help from AI writing tools. Owned by Daily Magic.
  • 2. How it serves the purpose: They make sure help is shown when words suggest someone may be in danger, and not shown for everyday phrases like "this deadline is killing me".
  • 3. Number of data points: About 100 sentences. The set grows as we add cases.
  • 4. Types of data points: Labelled "crisis" or "not crisis". Some are also marked as someone being hurt by another person.
  • 5. Copyright, trademark or patent; public domain: Daily Magic's own writing, protected by our copyright. Not public domain. No one else's copyrighted text, trademark or patent.
  • 6. Purchased or licensed: No. Neither purchased nor licensed.
  • 7. Personal information: No. Every sentence is invented. None are real people's words.
  • 8. Aggregate consumer information: No.
  • 9. Cleaning or processing: Some sentences are written on purpose with misspellings, curly apostrophes and extra spaces, to match how people really type.
  • 10. When the data was collected: From September 2026. Ongoing: we keep adding sentences.
  • 11. First used: September 2026.
  • 12. Synthetic data: Yes. Every sentence was written for testing; none were collected from people.

Behaviour checks

Example replies and expected results used to test Daily Magic's own checks around the Fairy Godmother: that she stays honest about what she remembers, that unkind wording is caught, and that she stays within her purpose.

  • 1. Sources or owners: Written for Daily Magic, with help from AI writing tools. Owned by Daily Magic.
  • 2. How it serves the purpose: They keep her honest and gentle, and keep the safety checks working as intended.
  • 3. Number of data points: Fewer than 100 test cases.
  • 4. Types of data points: Each case states the result we expect from the check.
  • 5. Copyright, trademark or patent; public domain: Daily Magic's own writing, protected by our copyright. Not public domain. No one else's copyrighted text, trademark or patent.
  • 6. Purchased or licensed: No. Neither purchased nor licensed.
  • 7. Personal information: No. Every example is invented.
  • 8. Aggregate consumer information: No.
  • 9. Cleaning or processing: None.
  • 10. When the data was collected: From September 2026. Ongoing.
  • 11. First used: September 2026.
  • 12. Synthetic data: Yes. Every example was written for testing.

Wellness knowledge notes

Written notes on general wellness topics, such as yoga, breathing and rest, that can be given to the AI model as background while it answers. They are not used to train any model.

  • 1. Sources or owners: Written for Daily Magic from general, publicly known knowledge, with help from AI writing tools. Owned by Daily Magic.
  • 2. How it serves the purpose: They help the Fairy Godmother give accurate, gentle, general wellness information.
  • 3. Number of data points: About 40 topic groups, roughly 2,800 lines of text.
  • 4. Types of data points: Unlabelled text, grouped by topic.
  • 5. Copyright, trademark or patent; public domain: Daily Magic's own writing, protected by our copyright. Not public domain. It mentions the names of some practices and schools, a few of which may be trademarks.
  • 6. Purchased or licensed: No. Neither purchased nor licensed.
  • 7. Personal information: It names some well-known teachers and authors when describing the history of a practice. It contains no information about Daily Magic users.
  • 8. Aggregate consumer information: No.
  • 9. Cleaning or processing: No outside data was cleaned or processed. The notes were written directly and edited as ordinary writing.
  • 10. When the data was collected: Written from at least June 2026. Ongoing: notes are updated.
  • 11. First used: At least since June 2026, our earliest record.
  • 12. Synthetic data: The text was drafted with help from AI writing tools. No other synthetic data was generated for it.

Fine-tuning

We have not fine-tuned any model. If we ever do, it will be only with the explicit, informed consent of everyone whose words would be used, and we will update this page before that version goes live.

What we will never do

We will never train, test or tune AI on your journal, your chats, your calendar, your tasks or anything else you keep in Daily Magic without asking you first, in plain words, with the choice to say no.

Updates and questions

We update this page before any new AI model, or any substantial change to one, goes live in Daily Magic.

Questions: [email protected].

Published under California Civil Code §3111 (the Generative Artificial Intelligence Training Data Transparency Act). Our crisis protocol