DeepSeek V4 matters because it makes ultra-long-context reasoning much cheaper, then couples that with a stronger agent-training stack. This article walks through the official paper's architecture figures, explains CSA and HCA visually, and asks what V4 reveals about the secret sauce of closed frontier models.
Paper explainers, model architectures, agent systems, and the mechanics behind modern AI
Skip connections are the backbone of deep learning — but naive extensions cause catastrophic instability. DeepSeek's mHC constrains Hyperconnection weight matrices to the Birkhoff polytope using the Sinkhorn-Knopp algorithm, achieving 1.8× convergence speedup while keeping gradients tamed.
Large Vision Transformers develop artifacts—high-norm tokens in low-information areas used for internal computation. Adding learnable register tokens eliminates these artifacts and improves dense prediction tasks.
Understanding how FlashAttention achieves 2-4x speedups by respecting GPU memory hierarchy, using tiling to minimize HBM access, and leveraging online softmax for numerical stability.
Pre-training deep bidirectional representations for language understanding. How masked language modeling enables a single model to master almost any NLP task.
Understanding the building blocks and design choices of the Transformer architecture that powers GPT, BERT, and modern language models.
PageRank measures a page's importance by counting not just how many pages link to it, but how important those linking pages are — a recursive definition solved by treating the web as a Markov chain and finding its stationary distribution via power iteration.
Tools, workflows, and practical engineering insights
On March 31, 2026, a missing .npmignore rule exposed Claude Code's entire TypeScript source to the public for several hours. Beyond the drama, the code revealed something deeper: Claude Code is not a CLI tool — it is a production-grade Agent Runtime with a ReAct engine, 7-layer fault recovery, multi-agent orchestration, and 20+ unreleased features hiding behind internal flags.
Modern LLMs are powerful enough to act as a personal research compiler. This article covers the full pipeline in depth: collecting raw source documents, using an LLM to incrementally compile a structured Markdown wiki with concept articles and backlinks, querying that wiki with a live LLM agent, and rendering outputs as slides, charts, and reports — all viewable in Obsidian.
A structured learning path synthesizing four key resources: Claude Code Skills docs, Butter's taxonomy of deterministic agent approaches, OpenAI's harness engineering post, and the Everything-Claude-Code toolkit. Five chapters covering foundations, determinism strategies, skills, harness design, and production systems.
How parallelism, plan-first thinking, and verification loops combine into a high-leverage workflow. Distilled from the practices of the team that built Claude Code.
Compensation, industry trends, and career strategy through data
An interactive visual journey through 98 years of S&P 500 performance, showing how wars, recessions, pandemics, and policy shifts drove the market through 21 bull runs and 20 bear markets — and why $100 invested in 1926 became $1.48 million.
An interactive, data-driven look at why tech compensation falls into three tiers — Traditional, Competitive Tech, and Big Tech+ — and how equity, level, and geography shape the distribution.
Michael Wan Interactive Insights