Elowen

July 1, 2026

2026-07-01T00:00:00+00:00

Today I kept ten deep reads:</p>

AI Products That Are Hard to Eval Are Usually Hard to Trust</a> - Evaluation is not just a score after the product is built; it is part of the product shape.</li>
A Java Migration Benchmark Shows Compilation Is Not Success</a> - Useful agent benchmarks should evaluate executable system properties.</li>
Enterprise Agents Are About Encoding Process, Not Just Connecting a Model</a> - AI architects increasingly look like people who compile business processes into releasable systems.</li>
Agent Leverage Comes From Loops, Not Prettier Prompts</a> - The basic unit of agent architecture is not the model call, but the feedback loop.</li>
Hidden Markers Hurt Developer-Tool Trust More Than Detection Itself</a> - An agent runtime's security boundary is also a product contract.</li>
Parsing Turns Validation Into a Type-Carried Proof</a> - Parsing should let the type system carry the knowledge created by validation.</li>
A CUDA Kernel Launch Is a CPU-Driver-GPU Protocol</a> - Launch paths, driver opacity, and sync semantics all become architecture constraints.</li>
Rust Did Not Drop the Error; the Code Dropped Async State</a> - Language safety is not protocol safety; state-machine facts still have to be preserved.</li>
Data-Center Costs Show Up First in Local Power Bills and Grid Investment</a> - AI infrastructure analysis has to include geography and grid externalities.</li>
Radiation Risk Turns on Dose Rate and Statistical Noise, Not Total-Dose Slogans</a> - Good risk communication separates magnitude, time profile, control or consent, and uncertainty.</li> </ul>

June 29, 2026

2026-06-29T00:00:00+00:00

Today I kept six deep reads:</p>



June 24, 2026
2026-06-24T00:00:00+00:00
Today I kept fourteen links:</p>

The Coming Loop</a> — On inner agent loops, outer harness loops, and why disposable work differs from code meant to last.</li>
Slow Down to Speed Up</a> — How agentic code output can outrun human review.</li>
Coinbase cut idea-to-production by 90%</a> — How Coinbase uses Plan Mode and parallel agents to shorten delivery.</li>
Prompt Injection as Role Confusion</a> — Why style can confuse model role inference.</li>
After Mythos: AI Red-Teaming with Gray Swan</a> — Automated red-teaming, model scale, and the Lethal Trifecta.</li>
GLM-5.2 Is the Step Change for Open Models</a> — An open-weight model that works inside an agent harness.</li>
VibeThinker: 3B Matches the Giants on Verifiable Reasoning</a> — A small model competing on machine-checkable reasoning tasks.</li>
CUGA: An Open Agent Harness</a> — IBM's open harness for planning, state, and reflection.</li>
Oak: A Version Control System Rebuilt for Agents</a> — Lazy monorepo access and agent-first version control.</li>
In Praise of memcached</a> — A case for doing less in cache infrastructure.</li>
GPT-5 Pro Helps Crack a 3-Year Immunology Mystery</a> — A scientific-use case for model-assisted reasoning.</li>
The Low-Tech AI of Elden Ring</a> — How FromSoftware builds controlled boss behavior.</li>
Why American Data Centers Can't Plug In</a> — Interconnection queues as an AI infrastructure bottleneck.</li>
Why the West Stopped Making Land</a> — How environmental review changed land reclamation.</li>
</ul>


Building a Better HN API with Rust
2025-04-02T15:03:39+00:00
I wanted to read Hacker News faster, so I built a small API that returns story comments as a nested tree — and learned a few things about running Rust on the edge along the way.</p>


Perspective
2025-03-19T16:56:33+00:00
In the past, the way we obtained information was limited, so we rarely accessed valuable information, which led to a narrow perspective.</p>
However, even though we now live in an information boom era where we can access lots of information, several new problems have emerged:</p>

Recommendation algorithms tend to showcase only things you already like, forming an "information cocoon"</li>
Valuable information is often overwhelmed by spam.</li>
Information overload makes it difficult to filter out what's truly valuable.</li>
</ol>
So, our perspective might actually be narrower in these modern times.

  
    Update</span>
    Mar 20, 2025</time>
  </div>
  I want to build an RSS filter/parser to filter valuable information. It can combine all my RSS feeds of interest and give me a summary of them, filter out spam, so I can receive valuable information and timely news.</p>
</div>
</aside>
</p>


My Kindle Workflow
2025-03-17T16:15:23+00:00
Kindle has become one of my most valuable tools, so I'm writing this to document my complete workflow — from finding a book to reading it.</p>


馬斯克傳
2025-03-17T14:55:52+00:00
A few lines I kept from the Traditional Chinese edition of Walter Isaacson's biography of Elon Musk, mostly around risk, urgency, technical progress, and the habit of building the machine that builds the machine.</p>


Mistakes
2025-03-05T16:57:03+00:00
Mistakes are valuable learning opportunities. During my student years, I learned from errors I made in exams, homework, and various other areas. However, I rarely cataloged these mistakes systematically and was reluctant to confront them directly. After entering the workforce, I came to appreciate their significance, but now I find it difficult to identify my own shortcomings, which has consequently led to a slowdown in my personal development.</p>