Yash Datta

A publication by Yash Datta

Read the Source.

Long-form, source-driven walkthroughs of the algorithms, numerical tricks, and engineering primitives that load-bear modern machine learning. New pieces by email at saucam.substack.com.

2026

  1. Aug 17 The sparsity that pays for everything else in DwarfStar LLM inference
  2. Jul 24 Long sessions without compaction AI agents
  3. Jun 23 Codeoid: cross-session memory and identity for Claude Code AI agents
  4. Jun 16 Three layers above your coding agents Agent security
  5. Jun 10 The rotation that rescues a 2-bit KV cache LLM inference
  6. May 26 Deploying MCP servers in production: the 2026 attack surface and the defense stack Agent security
  7. May 22 Whispers of Wealth, measured: 20 runs against the AP2 attack paper, on its own model Agent security
  8. May 21 What attention stores, and what DeepSeek changed Architecture
  9. May 20 MRV2: how vLLM rewrote its model runner, in six architectural moves LLM inference
  10. May 15 Streaming softmax: the recurrence that makes FlashAttention work Math foundations