Yuwei ZhangKarp's Anger, Figma's Squeeze, and the Clue Microsoft Office Left Behind
From Karp's charge against model vendors, to Figma's first-draft problem, to the lesson Microsoft Office already left behind.
The Role of Markdown Files in the Harness
Small markdown files do not change model weights, but when and how the harness injects them decides whether they act as constraints or as noise.
What Makes a Good Researcher
Insight is the interest earned on systematic knowledge; framing fixes the ceiling; pitching debugs your understanding; interest funds the whole loop.
skVM: Building a Cross-Platform Compiler and Runtime for Agent Skills
Why skills break when they change machines, and how SkVM borrows the compiler/VM playbook to make them portable, verifiable, and self-optimizing.
Decoding Claude Code's Memory System Design
Markdown over vector databases, model-driven recall over embeddings, and offline consolidation: the design bets inside Claude Code's memory, and why they hold up.
Harness Engineering and Agent Capability Orchestration
Same model, +13.7 points on Terminal Bench: what prompt and context engineering miss, and why the next layer of agent reliability is built outside the model.
MemEvolve: Dynamic Structural Evolution of Agent Memory
What happens when the memory architecture, not just the memory strategy, becomes searchable, and what the three evolved systems reveal about the trade-off surface.
From Benchmark to Alignment
Facts are universal; meaning is personal. Why memory is the infrastructure that turns capable models into aligned ones.
From RLHF to RL'ME'F
Rather than predicting the next token, predict your next thought: how Me.bot bets on parameterized personal memory over RAG.