§ 04/ Writing
Medium2026, 7 min
The Fastest Way to Speed Up an AI Model? Make It Skip Parts of Its Own Brain
On skip-attention, sliding windows, and why on-device speed comes from measurement, not theory.
Here is a weird idea. If you want an AI model on your phone to respond faster, one of the best things you can do is have it skip steps entirely in some of its layers. On TTFT, sliding windows, skip attention, and letting the device tell you what actually wins.
Read on Medium
Medium2025, 4 min
Understanding the Self-Attention Mechanism in Transformers (with PyTorch Code)
Q, K, V from scratch. What the softmax is actually softmax-ing over.
A step by step walk through self-attention with runnable PyTorch. Query, Key, Value; the scaled dot-product; softmax normalization; masking; multi-head. If BERT and GPT feel like magic to you, this is where the magic actually happens.
Read on Medium