Machine-learning

How X ranks candidates using a tiny Grok-based transformer with candidate isolation masking — scoring all posts in a single forward pass without them interfering with each other.

How X finds ~200 candidate posts from billions in milliseconds — Thunder keeps in-network posts in memory, while Phoenix Retrieval uses a two-tower transformer to discover out-of-network content via embedding similarity search.

X open-sourced their For You feed recommendation algorithm. This is Part 1 of a deep dive into how it works — from the four-codebase architecture to the full pipeline overview, and the one surprising design choice that changes everything.

I argue that right now AI is not a replacement, but an enabler for people to move fast and breakthrough. You utilize the power of LLM that is reasoning, planning and contain general knowledge.

Exploring how to incorporate side information into graph embedding for cold start recommendation problems using MovieLens dataset and PyTorch.