Technical Logs
Thoughts on ML Engineering, Systems, and Code.
Jan 2026
A New Axis of Sparsity for Large Language Models
DeepSeek's Engram introduces conditional memory to LLMs - a revolutionary approach that separates knowledge retrieval from reasoning, achieving massive performance gains.
Jan 2026DilatedConvBlock: When Convolutions Learn to Breathe
Exploring the mechanics of dilated convolutions and their impact on receptive fields in deep learning models.