Masaki Mu
In the middle of the journey, I came to myself.
Latest
-
Humans as Weak Supervisors: What AAR Reveals About Alignment
Anthropic's AAR project is ostensibly about autonomous AI research. Viewed differently, it's a meta-validation of weak-to-strong alignment: humans as weak supervisors, wielding evaluation environment design as their last leverage point over models that exceed their own capabilities.
-
Brain ≠ Hands: Dissecting Anthropic's Managed Agents Architecture
Anthropic published their Managed Agents architecture. Under the hood, the real insight isn't the three-layer split — it's two counterintuitive decouplings: Session ≠ Context, and Tool execution doesn't live next to the Agent.
-
The Agent Trapped in the Same River
A month of building memory for an AI Agent. What we learned wasn't how to build a pipeline — it was the fundamental difference between a database and memory: databases preserve facts, memory lets knowledge grow.
-
Agent Context Compaction: From Cursor's Self-Compaction to Claude Code and Codex CLI
Cursor trains models to self-compact via RL for long-horizon tasks. We dissect the KV cache reuse mechanism and compare it with Claude Code's 9-section structured compact and Codex CLI's auto-compact through reverse engineering.
-
Bub Tape Architecture Deep Dive: From Traceable Memory to Controllable Context Windows
A deep breakdown of Bub's Tape subsystem: storage structure, context selection, handoff/reset semantics, search mechanism, and engineering trade-offs.