(Note: This model is for Trial Testing - Underfed until completion for base model fully hand crafted with 600k without scraping)

Fatty AI Singapore – Beta Trial Testing Version 1.6 – Supervised Machine Learning Model

A 42M Parameter Singapore Soul in the Making. Local. Clean. From Scratch. – Build by Singaporean

The Fatty AI Method: Pure Supervised Learning
Fatty AI is built from the ground up using Supervised Machine Learning. This is the core of how it thinks, but it’s a very different process from what you might have heard from large AI companies.

Fatty will only expand tools after the base is stable with more learning.

What is Supervised Learning?
In simple terms, Supervised Learning is like teaching a child with flashcards.

You show the child a flashcard with a Question (the input).

On the back is the correct Answer (the desired output).

By studying thousands of these flashcards, the child learns the underlying patterns and eventually can answer new, similar questions on their own.

This is exactly how Fatty AI was trained.

How Fatty AI Was Supervised?

I acted as the teacher. For over 600,000 examples, approximates around 786 million characters in this version which is still underfed. I provided the model with a perfect pair:-

A precise instruction or question ( e.g., How do I walk from Concorde Hotel to Camden Medical? )

The exact, correct answer I wanted it to learn (e.g., Starting point: Concorde Hotel… [ – detailed route – ]…)

The model’s only job was to analyze these millions of Question and Answer, pairs and discover the deep patterns that connect them. Through this intense, focused training over hundreds of epochs, it didn’t just memorize, it learned to understand the structure of navigation questions, the properties of locations, and how to generate a correct response.

The Key Difference: Deep Education vs Surface Level Manners
This is where Fatty AI differs completely from many models.

Fatty AI – Deep Education

Step 1: Foundation Learns from my clean, hand written prompts. Its entire worldview is built from high quality, structured knowledge.

Step 2: Supervised Phasem, This is its entire life. Supervision is the primary, continuous process that builds its core intelligence.

Result:- A specialist with a pure and reliable knowledge base, capable of true reasoning within its domain.

Like many solo developers, whether fine-tuning or building a true base model from absolute scratch, the early days are pure torture. When my training data was only 2,000 to 3,000 lines, every epoch felt pointless. I wanted to give up a thousand times. But giving up was never an option. So I kept pushing, day after day, night after night, since July 2025, Fatty’s actual birthday month. Yes, you read that right: this is a genuine solo built base model, hand fed by one stubborn Singaporean who refused to quit.

What is Fatty AI?

Fatty AI is an ambitious experiment: a compact base model built entirely from scratch. No pre-training, no scraping. No Tag, Intent training or place holder. Every piece of knowledge was carefully hand fed, not web scraped, ensuring zero copyright risks. It’s designed to run locally on Android phones ( FP32, from 4GB to 16gb+6gb RAM ), with the goal of becoming a true Singapore expert in your pocket.

Our Approach: A Different Path

The common belief is that advanced AI requires immense resources. We’re exploring a different path. By prioritizing relentless focus and over 600,000 hand crafted lessons over financial capital, Fatty AI is our testbed for a more efficient AI paradigm.

While conventional wisdom relies on massive data centers, we chose depth over breadth. Instead of scraping the entire internet, we wrote those 600,000+ precise lessons estimated around 786 million characters. The result is a compact model that aims to prove focused, efficient, and capable AI is a viable path forward.

For more info: https://fatty.com.sg
suggestion: [email protected]