A POLYGON DAYDREAMMake weird things.
↓Make weird things.
Stay curious.
Scroll to leave the rainbow trail and enter the cat wall.
Recommended
Released Two Astronomy Apps
A 360° Product Rotation System for Web Shops Using MinimaxH3
All 61 Episodes of Our Space Documentary Are Now Available in Three Languages
Complete 60-Episode, Three-Language Playlists for The Art of War Now Available
Sleep Better! Relaxing Algorithm Stories | 3 Languages, All 60 Episodes
My Son Said He Wanted to Watch History, So I Created a Video Series About Life 30,000 Years Ago
I Built a Croatian Text-to-Speech Model with 32 Million Parameters: Comparing Voices During Training
Following my Croatian RAG LLM, I decided to build a small text-to-speech (TTS) model specifically for Croatian. The goal was real-time performance on a smartphone CPU. I used distillation: I had a large TTS model (HiggsAudio v3) read Croatian scripts to create training data, then trained a small model from scratch on t…
Reports for October 25, 2026
Testing SLM RAG Performance: Comparing 25 1B–9B Models in Croatian and English
Having built a Croatian RAG model and even quantized it, I next looked into how usable the small models (SLMs) currently available are for RAG. I tested 25 models—10 in the 1B class, 8 in the 3B class, and 7 in the 7B class—plus my own 320M model, on the same questions in Croatian and English. Update (same day): I also…
Reports for October 24, 2026
I Quantized My 300M-Parameter LLM and Compared CPU Speed and Accuracy
I took the 320M-parameter model for Croatian RAG that I built in the previous article. Since it is meant to run on a CPU, this time I quantized it to see how much smaller and faster it could get—and how much accuracy it would lose. Here is the conclusion first. - Q6K hits the sweet spot. It is one-fifth the size (254MB…
I Built a Croatian RAG LLM from Scratch with 300M Parameters—It Runs on a CPU
Following the previous GPT-2, this time I built a Croatian LLM from scratch on a single RTX 4090. The goal was RAG: a model that finds answers within supplied text and responds while quoting one supporting sentence. And it is intended to run in production on a CPU. Here is the conclusion first. - About 68 GPU hours fro…