Thoughts on ML models, AI applications, distributed systems, and data platforms.
What studying Claude Code's architecture taught us — how the agent loop, context management, tool schema cost, sub-agent isolation, coordination protocol, and reliability infrastructure actually work in production.
How to write skills that trigger reliably, load efficiently, and stay maintainable, with concrete examples from official docs, the open specification, and public skill repositories.
A deep dive into RL infrastructure for LLMs, from RLHF to RLVR to Agentic RL, with head-to-head experiments comparing rule-based vs. LLM-as-a-Judge rewards and sync vs. async training on SkyRL.
How independently-built agent systems converged on the same architecture for managing finite context windows.
How functional architecture helps with durability, security, scalability, and maintainability in agent systems.
From PPO to GRPO: tracing the evolution of policy optimization algorithms for efficient LLM alignment and comparing their real-world applications.
From collaborative filtering to generative AI: tracing the evolution of recommender systems through the lens of the bitter lesson
How declarative ML creates a shared language for human-agent collaboration.
A comprehensive analysis of DeepSeek V3 and R1 models, covering key innovations like MLA, MoE, MTP, GRPO, and the evolution from base models to reasoning-capable systems.
Get notified when we publish new articles about ML models, AI applications, and data platforms.