A place for writing and making
Atelier
Notes in progress. Work made with care.
Thoughts take shape in the Blog; work ready to share, together with its notes and files, lives in Articrafts.
- Writing
- Questions, arguments, and field notes, worked through in public.
- Making
- Papers, code, reports, and experiments, gathered here as they are ready and open to later revision.
- Recurring threads
- Methods and subjects that return across different pieces.
Recently
New writing and work
Recent additions from the Blog and Articrafts.
Writing
Recent writing
-
The Death of Code Review: Toward Agent-Legible Systems
Coding agents are reshaping the cost model of software engineering: code generation is becoming increasingly cheap, while human attention is becoming the new serial bottleneck. Software trustworthiness should not depend on the authority of an author or reviewer, but arise from verifiable and falsifiable system mechanisms. This essay proposes Agent-Legible Systems: systems designed from the outset so that agents can query, experiment with, and verify runtime state, critical invariants, and causal evidence, forming an automated verification loop through type systems, testing, CI, profiling, tracing, and more. Human judgment then moves upward from line-by-line implementation review to domain models, system contracts, and irreversible architectural decisions.
-
Local Optima Under Closed Conditions
Array-based red-black trees, gaokao finale problems, and PhDs who cannot find jobs are, at bottom, the same thing: closed systems can select for local optima, but the definition of this 'optimum' is itself distorted by the closed conditions. The information environment determines the output efficiency of intelligence, and you cannot price what lies outside your cognition.
-
AI as a New Infrastructure for Knowledge Transmission
The rise in humanity's average capability comes less from individuals studying longer than from the continuous upgrading of how knowledge is diffused, compressed, and invoked.
Work
Recent work
-
From Feature Checklist to Evolvable System: AI Job Workspace Product and Architecture Internship Report
A five-page internship report on AI Job Workspace product design, domain modeling, API contracts, observability, and AI-native team collaboration.
-
Cinder: Design and Implementation of a Compact CUDA Tensor Library
A design record for a compact Python/C++/CUDA dense tensor library, covering tensor semantics, shape planning, device-memory ownership, kernel protocols, and extension risks.
-
Before MPI: Locality and Kernel Policy in One-Sided Jacobi SVD
A 420-configuration factorial study of one-sided Jacobi SVD across six matrix layouts, five kernel policies, and fourteen matrix families, separating locality, convergence, and parallel-wrapper overhead.