- formatting
- images
- links
- math
- code
- blockquotes
- external-services
•
•
•
•
•
•
-
LoRA and QLoRA: Fine-Tuning 70 B Models on a Consumer GPU
LoRA, QLoRA, and the PEFT ecosystem — how the intrinsic dimensionality hypothesis lets us fine-tune billion-parameter models on a single GPU.
-
RoPE and ALiBi: Giving Transformers Unlimited Memory
How RoPE, ALiBi, and YaRN enable language models to handle context windows from 4 k to over 1 million tokens.
-
Vision Transformers: How Attention Conquered Computer Vision
From patch embeddings to DINOv2 — the complete story of how Transformers revolutionized computer vision.
-
Diffusion Models: The Probabilistic Engine Behind Generative AI
A rigorous but accessible walkthrough of DDPM, score matching, and latent diffusion — the mathematical backbone of Stable Diffusion and DALL·E.
-
RLHF and DPO: Teaching Language Models to Be Helpful and Harmless
The complete alignment pipeline — from SFT to RLHF with PPO, to Direct Preference Optimization that eliminates the reward model entirely.