Posts

Hello, this is DataFox. If you’ve visited datafox.tw recently, you might have noticed something – I haven’t updated any new articles since June 12th. Many people have also asked me: “Didn’t you say you’d write course reviews?” “What about AI news?” “Why no new articles recently?” To be honest, it’s not that I lack ideas, but rather that I’ve been truly too busy. During this time, in addition to my Google internship, I’ve also been simultaneously juggling three papers (including AAAI and ICAIF), industry-academia collaboration projects, and various research tasks for grad school. I don’t know why, but things that were originally spread out over different times all ended up crammed into the same month or so. Furthermore, I still hope to maintain a normal life: meeting friends, exercising, sleeping, and not completely burning myself out. Consequently, managing my self-media naturally took a back seat. ...

Hello, this is DataFox. If you’ve recently visited datafox.tw, you might have noticed something – I haven’t updated any new articles since June 12th. Many people have also asked me: “Didn’t you say you were going to write course reviews?”, “What about new AI insights?”, “Why no new articles recently?” Honestly, it’s not that I lack ideas, but I’m truly too busy. During this period, in addition to my Google internship, I’ve also been simultaneously juggling three papers (including AAAI and ICAIF), industry-academia collaboration projects, and various research tasks for my graduate studies. I don’t know why, but things that were originally spread out over different times all ended up piling up within the same month or so. Besides all this, I still want to maintain a normal life: seeing friends, exercising, sleeping, and not completely burning myself out. So, naturally, managing my personal media presence took a back seat. ...

最近跟朋友聊求職，發現大家都被一種荒謬的焦慮感籠罩。現在打開社群，滿地都是「AI 讓你十倍速開發」、「三分鐘寫完一個 APP」的東西。我自己也變得越來越倚賴純 vibe coding。 在 2024 以前，我要開發一個功能，還會想到底要貼哪一段 code 給 GPT，讓 GPT 回說這部分要怎麼處理。那個時候，我至少有動腦去找到問題根源，只不過解決它的方式比較粗糙。 但自從我去年七月從歐洲回到台北，八月下載 Cursor 來用，到現在 Claude Code、Antigravity、Codex 百家爭鳴，我們漸漸開始發現： 現在工程師的功用，好像真的正在指數級銳減。 但與此同時，有件事情也同時發生了。 在求職的路上，或者是任何需要展示出自己比別人優秀的路上，我們發現：工時沒有變短，生活品質也沒有變好。反而是整個產業界對「人」的期待值，跟著這些宣傳語同幅度、甚至超額通膨了。 以前一個人花兩三個月打磨出一個有模有樣的 side project，大家會覺得你底子不錯。 現在就會變成，面試官看著你手上的專案，心裡的 OS 可能是： 「這大概是用 AI 兩三天拼出來的吧，沒什麼了不起。」 然後連我自己都開始懷疑。 在這種通膨下，我們好像被逼著得是十項全能的超人，才有資格在市場上競爭。 廉價指標飽和之後，證照和成績還剩多少鑑別度？ 然後大家開始瘋狂去刷證照、刷成績。 但說穿了，在這個連小學生都能用 AI 輔助備考的年代，那些有題庫、靠死記硬背的「世俗標章」，早就貶值到不行了。 當所有人都拿著滿分證照去面試時，這就是一個廉價指標飽和的死胡同。 那到底要怎麼證明自己不是一個只會發送 prompt 的「AI 打字員」，而是能帶來十倍效益的人？ 真正清醒的面試作業：Prototype 只是基本，解剖報告才是重點 前陣子我看到一份極度清醒的面試作業要求，簡直是直接賞了那些整天吹捧 vibe coding 的人一巴掌。 那家公司劈頭就說： Working Prototype 是基本，但他們不要一個只是包著 LLM 外皮的玩具。 他們真正要看的是： PRD（產品需求文件） TDD（技術設計文件） 以及三個靈魂拷問： What I personally built What I reused What broke and how I debugged it ...

Lately, my life has been incredibly hectic. While desperately trying to supplement data for my thesis, a practical collaboration project on the other side also presented a very real and extremely tormenting engineering challenge. This article won’t delve into too much specific business logic; instead, it will simply discuss a huge performance pitfall we stumbled upon in our system architecture, and what absurd stress-relief behaviors an engineer might resort to when stuck debugging. ...

Readers who have been following this website might have noticed the site undergoing a severe “labor pain” period within the last few hours — meaning the screen went completely blank and crashed when switching languages. But after some emergency fixes and a major overhaul, I am delighted to announce: this website has officially upgraded to a “Native Bilingual Architecture (Bilingual Native)”! Ditching the Hacky CSS Hiding Method In the previous article, I mentioned using AI to automatically translate articles. The approach at the time was quite “hacky”: I had a Python script embed the translated English along with the original Chinese into the same Markdown file, and then used custom buttons and CSS display: none to forcibly hide or show the corresponding language sections. ...

During the process of building this personal website, I’ve always faced a dilemma: I want to record and create in my most familiar and heartfelt native language (Traditional Chinese), but at the same time, I also aspire to connect with the world, allowing more English native speakers or international readers interested in my technical articles to discover the potential of this content. Furthermore, from a technical blog management perspective, bilingual content significantly aids search engine optimization (SEO), bringing a wider range of organic traffic to the website. ...

From the ‘Tape-Finding’ Mystique of Geoguessr to the Shortcut Learning Myth in Deep Learning: What Happens When We Over-rely on Certain Features? 1. The Days of Being Dropped in the Middle of Nowhere: The Romance of Vibe Guessing Ever since my exchange student days ended, I’ve unexpectedly become captivated by Geoguessr. Its charm lies in not needing to download any app; you can simply open your browser and be ‘randomly airdropped’ anywhere in the world, anytime. While playing, I not only revisit street views I’ve walked through but also discovered I was gradually developing a superpower – ‘Vibe guessing’. ...

Recently, while reviewing the backend data for datafox.tw, I discovered a very harsh reality: some of the hardcore technical articles I spent a great deal of time writing were virtually invisible to search engines and AI crawlers. In 2026, an era where AI SEO and GEO (Generative Engine Optimization) dominate attention, my failure to focus on EEAT was a critical strategic mistake. 🔍 The Fundamental Problem: I’m Trying to Compete with the World’s Strongest Giants for the Same Keywords I previously wrote an article discussing flash vs thinking kv cache test time compute. I thought this article, being highly technical, would attract decent organic traffic. The result? The data was dead silent. ...

As I delved deeper into LLM applications and Agent systems, I became utterly disgusted by the industry’s “AI sycophant” style of mystical marketing. Vendors now constantly boast that their new models “can think” and “have logic,” as if adding a “Pro” or “Thinking” suffix suddenly makes a neural network sprout a human brain. If we don’t adopt an attitude of skepticism and questioning, we can easily be led astray by these attractive claims. Today, let’s coldly dissect the true differences between Flash models and Thinking models, and what those so-called “thinking processes” actually entail, starting from their underlying operational mechanisms and physical limitations. ...

At the time, I used to think that a Decoder was essentially just playing “word-chain” based on the input, unable to pre-determine what it was going to generate as a whole. But looking back now, if we don’t blindly join the hype of those AI sycophants, spouting pretty phrases like “because GPT is more spiritual,” but instead dispassionately dissect it from the foundational layers of Information Theory and engineering reality, you’ll find that the triumph of Decoder-only architectures is, fundamentally, an inevitability dictated by mathematical and physical constraints. ...