Yash Datta

Yash Datta — Read the Source

Read the Source takes one mechanism behind modern AI at a time and works it all the way down: the theory in full, a runnable implementation, the safety and alignment properties, and what it takes to run it in production. Reading one of these should leave you understanding the mechanism: why it works, and where it breaks.

I work at the intersection of AI and cybersecurity, building AI safety models and agent control fabric at Highflame. At vibz.art I build Sanskrit-first AI infrastructure for Indic knowledge corpora. Agent Almanac is my publication of recurring, rigorous benchmarks of AI agents on real-world tasks.

Featured piece

The sparsity that pays for everything else in DwarfStar

DeepSeek V4 Flash keeps 97.5% of its parameters in routed experts that fire 2.34% of the time. Five design decisions in DwarfStar, the DeepSeek V4 engine from antirez, all collect on that one gap.

Read the full article →

Read the Source

All writing →
  1. Aug 17, 2026 The sparsity that pays for everything else in DwarfStar DeepSeek V4 Flash keeps 97.5% of its parameters in routed experts that fire 2.34% of the time. Five design decisions in DwarfStar, the DeepSeek V4 engine from antirez, all collect on that one gap. LLM inference
  2. Jul 24, 2026 Long sessions without compaction How codeoid keeps month-long agent sessions inside a small context window: rotate the window instead of summarizing it, page verbatim history back through MCP tools, and fork the same conversation onto a different backend. AI agents
  3. Jun 23, 2026 Codeoid: cross-session memory and identity for Claude Code An open-source Bun daemon that adds verbatim cross-session memory, three-layer context reduction, and per-agent ZeroID identity to Claude Code. Plus codeoid-ui: a native Rust/Ratatui cockpit. AI agents
  4. Jun 16, 2026 Three layers above your coding agents A map of the layer forming on top of Claude Code and Codex (single harnesses, parallel-run UIs, and the meta-harness), and why governance, not orchestration, is the part that decides whether a company can adopt any of it. Agent security

Papers

  1. JavelinGuard: Low-Cost Transformer Architectures for LLM Security Yash Datta, Sharath Rajasekar · arXiv, 2025

    First author. Low-cost transformer architectures for detecting malicious intent in LLM interactions.

  2. DeepContext: Stateful Real-Time Detection of Multi-Turn Adversarial Intent Drift in LLMs Justin Albrethsen, Yash Datta, Kunal Kumar, Sharath Rajasekar · arXiv, 2026

    Originated the core idea; led and executed by Justin Albrethsen.

New pieces by email — saucam.substack.com.