Personal blog of Poga - transforming complex problems into fast feedback loops
# systemd & AI Let's put the two most hated things in tech together and start solving problems. ## I don't need to productize everything before I start solving problems In this age, people rush to productize everything. They want to build a SaaS, a startup, a company. They want to make money from us. Everyone wants to be a thought leader, a solopreneur, a billionaire. But I want to solve problems.…
# References for Agent Design - [https://utkarshkanwat.com/writing/betting-against-agents/](https://utkarshkanwat.com/writing/betting-against-agents/) - [ttps://userjot.com/blog/best-practices-building-agentic-ai-systems](https://userjot.com/blog/best-practices-building-agentic-ai-systems) - [https://kirshatrov.com/posts/claude-code-internals](https://kirshatrov.com/posts/claude-code-internals) -…
# Runtime over Syntax We need runtimes that help humans and LLMs solve problems together. LLMs read code like water flows through paths. Give them clear code and they understand fast. Complex tricks slow them down. Simple wins. But reading code is just the beginning. The real partnership happens in the feedback loop: LLM writes, tools respond, human guides, LLM learns. Static checkers become…
# Recently: 2025 06 18 ## Body I started weight training this year and have been making steady progress. ## Reading [The Who Care Era](https://dansinker.com/posts/2025-05-23-who-cares/) is probably the most important article I've read this year. > As the culture of the Who Cares Era grinds towards the lowest common denominator, support those that are making real things. Listen to something with…
# Breaking The Spell 打破魔咒 Translated from [Phrack Issue 71](https://phrack.org/issues/71/1). --- It can feel like the world is in a dreamlike state; a hype-driven delirium, fueled by venture capital and the promises of untold riches and influence. Everyone seems to be rushing to implement the latest thing, hoping to find a magic bullet to solve problems they may not have, or even understand. While…
# Observability Checklist Catch issues before users do 1. Have I defined clear "broken" conditions that truly reflect what users experience as failure? 2. How will I be notified when my service degrades or fails? Is this notification faster than user reports? 3. What leading indicators might show trouble before a full outage occurs? Am I monitoring these? 4. Have I implemented smoke tests that…
# Music as Precision Medicine  I make music in a unique way - limited to about two hours of sound exposure daily before headaches set in. Those two hours are my precision medicine - carefully timed, perfectly dosed, and specifically formulated for maximum effect. I plan each session like a curated tour through a specific slice of music culture. One day might be dedicated to…
# Code as Evolving Scaffolding Every program is just scaffolding for your next understanding. Been thinking about this scaffolding metaphor. It works on multiple levels: Immediate level: - Code = temporary structure - Real value = data collected + the collected test suites - Don't over-invest in the scaffolding Deeper level: Each program is scaffolding for understanding the next version of the…
# Writing directly for the web browers The blog you're reading is written directly for web browsers. no extra build step. The blog have many reincarnations, the longest lived one is based on [hugo](https://gohugo.io/). However, even a static site generator will deprecate some features one day. and my words need to follow the convention of that static site generator. And one day, the blog is…
Recently I incorporated a new technique into my workflow: using inline-snapshot-tests as both a REPL and a documentation tool. The idea is summarized in: 1. [What if writing tests was a joyful experience?](https://blog.janestreet.com/the-joy-of-expect-tests/) 2. [My Kind of REPL](https://ianthehenry.com/posts/my-kind-of-repl/) Basically you write a test that inline encodes the expected output of a…
1. Markets can only be efficient with perfect information, but perfect information is impossible in practice. 2. Markets can only be efficient when participants have complete freedom to make decisions, but this freedom is always constrained by real-world limitations. 3. Therefore: Truly efficient markets don't exist. 4. To survive in these inefficient markets, companies create boundaries that…
My take on AI: 20% of my skills are now obsolete, but the other 80% just got way more valuable. ## Why? 1. Engineering is all about trade-offs. If someone can clearly define all the trade-offs in detail, they can probably code it themselves. 2. Engineering exists because humans suck at precisely describing what they want and how to build it. 3. My real job is helping people figure out what they…
Here's my note on doing the [One billion rows challenge](https://github.com/gunnarmorling/1brc). My goal here is to: 1. find the reasonably optimized solution that's still readable and maintainable for a new-grade programmer. 2. Show that how easy it is to optimize code with the help of the flamegraph. You can find the code [here](https://github.com/poga/onebrc/blob/main/src/main.rs) # Step 1.…
Under the surveillance capitalism standard, [g0v.social](https://g0v.social) is a strange little social network. Traditionally, social networks seek to radicalize people; the more toxic you become, the more traffic the platform will receive, and thus the greater the profit. We used to use the internet to escape from reality. But now we return to reality to escape the toxicity of the internet. The…
# Cheap Complexity and Cybersecurity A recurrent theme of modern security is the importance of complexity and the difficulty of containing it. With GPUs, NICs, basebands, and specialized hardware such as TPUs, things are objectively getting more complicated at a superlinear rate. In a world where complexity continues to grow, it is hard to provide security guarantees. The problem of escalating…
**Context**: Machine Learning Projects are mainly just a complex interdependent pipeline. We desperately need a better abstraction for them. ## Hidden Complexity Here's a simple rule I used to evaluate if a problem has a well-defined programmable abstraction: can I implement it with just one text file and a text editor? - With [Infrastructure as…
[Using AntiPatterns to avoid MLOps Mistakes](https://arxiv.org/abs/2107.00079), Nikil Muralidhar et. al. I learned about this survey paper from [The Data Exchange](https://thedataexchange.media/mlops-anti-patterns/) podcast. It's a good introduction for people who just started deploying their machine learning project to production. The paper focused on _supervised learning_, and _forcasting…
 **這是 COSCUP 2016 的講題逐字稿，投影片在[此](http://www.slideshare.net/poga/log-65475572)** -- 自古以來，軟體工程師都在追求好維護，容易理解的軟體架構。傳統上，我們需要參與過各種大型軟體專案，從中獲取經驗，或是透過昂貴的課程，大量的論文，才能從前人的經驗中學到一些方法。 今天，我想試著透過 LOG 這個人人都碰過的資料結構，來解釋許多複雜系統的核心，只要你能理解 LOG，你就能設計出好理解、容易維護的系統架構。 -- 什麼是 LOG？每個人第一次寫程式時，輸出的 “Hello World” 是一種 LOG，工作時使用 Slack 有 LOG，你的伺服器有 Access Log。LOG 就是由兩個特性組成的資料結構： * 訊息按照時序出現 * 出現過的訊息不會改變…