Self Invented Reasoning – Singapore AI Fatty – Artificial Intelligence 42M Model’s Growth

Model Reasoning with Self Invented Placeholders – Singapore AI Fatty

Fatty is a 42M parameter Singapore AI model trained from scratch on 600,000 hand written. Example of Singapore walking routes.
From seeing similar route patterns repeated thousands of times, the model spontaneously invented its own placeholder system:

Hilton hotel ->  instantly becomes Hilton Singapore Orchard, 333 Orchard Road

A note from real testing – During daily testing, we often see responses like the Hilton example above – the model was never trained on the single word Hilton as a standalone query. When the user types only Hilton or Hilton hotel, Fatty grabs the name from the user input, expands it with the full hotel details he learned from route prompts, and inserts it into both the starting point and ending point of a walking route.

This creates some funny (and sometimes wrong) replies when start = end, but it perfectly demonstrates the core behaviour: the model has quietly built its own placeholder system from seeing thousands of similar route patterns.

We keep these examples public because they are 100 % real – no cherry picking, no hidden prompts. They show both the magic and the growing pains of a tiny model that is still learning every day.

It then plugs the expanded name into the walking template it also invented itself.

PS: Some may consider this as hallucination, however Fatty is underfed and personally i think this is emergent intelligence with zero tricks and 100% clean data. Without need of Rag, function calling, entity linking, Json schemas.

Alright, same like 4D, I am just an observer though I didn’t buy any 4D 🤣 This image is about Fatty, not about the sudden Popup of 4D to let me witness the lucky number for 30-11-2025 (Saturday).

What’s this image about? It is another example without the training example for Court Nojima as I missed it during the typing of Orchard Road routing training examples. During the process, I have added “The Heeren Shopping Centre or Court Nojima” into multiple similar routing for Fatty AI Model. However there is no trace of any training example for Where is Court Nojima. This is another interesting emergent reasoning example for Fatty.

However, as the creator for Fatty, whenever I notice gaps like this, I add new training samples to confirm and strengthen the model’s reasoning.

Important note:

The placeholder expansion itself is remarkably accurate.
The generated walking route is sometimes still wrong or incomplete because the model is tiny and training is ongoing.
This behavior emerged naturally – no entity lists, no tags, no external rules.
Just pure pattern density from repetitive, hand written data.
We show it because it proves even a 42M artificial intelligence model can invent its own reasoning shortcuts when fed clean, focused data for long enough.

Fatty estimated release date will be in 2026. Dec 2025 version is a Trial Testing Beta. – Click Here Back To Main Fatty Information Page

Fatty’s account page on Github -> https://github.com/YiPinQian/Fatty

Quote of the day on 4th December 2025 – If you can’t see a problem when it seems too perfect, you can’t fix anything at all.

Local or Foreign Users Suggestion or Ideas for natural hand fed data please send us an Email: Fatty Model Email Address