- formatting
- images
- links
- math
- code
- blockquotes
- external-services
•
•
•
•
•
•
-
Graph Neural Networks and Foundation Models for Science
How GNNs and graph-aware Transformers are enabling breakthroughs in drug discovery, materials science, and protein structure prediction.
-
Contrastive Self-Supervised Learning: CLIP, SimCLR, and DINO
SimCLR, MoCo, BYOL, and DINO — the elegant mathematics of learning powerful representations by contrasting augmented views, without any labels.
-
The Transformer Architecture: A First-Principles Deep Dive
A rigorous technical walkthrough of every sublayer in the original Transformer — the architecture underpinning virtually all modern AI.
-
Mechanistic Interpretability: Reverse-Engineering the Transformer
How researchers use circuits, activation patching, and the logit lens to understand exactly what computations happen inside Transformer models.
-
Speculative Decoding: 3× Faster LLM Inference for Free
How speculative decoding uses a small draft model and one parallel verification pass to dramatically accelerate autoregressive inference.