Posts

「If we’re on the path to the Turing machine, Seq2Seq itself makes more sense than a Decoder." This is a sentence I wrote in my notes in 2022. Back then, I was experimenting with an early version of GPT-2, and I kept wondering: What exactly is this thing? In comparison, models like BART or T5, based on the Seq2Seq concept, seemed much more reasonable. Unexpectedly, a few years later, in this AI arms race, it’s the Decoder-only architecture that has claimed the MVP title. ...

In “Post-Completion Notes for datafox,” I proudly introduced how I used JSON-LD for AI-SEO, how I embedded Person Schema into extend_head.html to help Perplexity recognize me, and how I made AI search engines understand my technical boundaries. At the time, I thought I was doing something quite remarkable. Approximately two months later, I opened Google Analytics and discovered a problem. My traffic almost entirely came from my own promotion. Every time I posted → shared on LinkedIn → shared in groups → traffic would spike briefly → then disappear. ...

Why I wanted to do this There’s a common misconception about static websites, which is that “static” means dead and unchangeable. But when combined with CI/CD, they can achieve things many people wouldn’t expect. I wanted to do one thing: to automatically update a certain field on this website every week, without me having to manually write anything. My motivations were twofold: The first was purely to test how far static deployments like GitHub Pages can go with automation. From data scraping, calling LLMs, to finally committing to the repo to trigger deployment, the entire chain can be completed within GitHub Actions, without any backend servers. This in itself is very interesting. ...

🚀 The Price of #VibeCoding: It Cost Me 6 Gmail Accounts to Build This RAG Platform For our AI club course, I wanted to use v0 (the AI coding platform by Vercel) to build a service that could be deployed seamlessly. My initial thought was: “If I can do it, my students definitely can too.” I was wrong. Because my early prompts weren’t sharp enough—and because of v0’s inherent “intelligence gaps” (it often misses details or completely fails to grasp the spec, leading to hallucinations)—I burned through account after account without ever finishing within the $5 free quota. ...

Although I didn’t directly serve as a lecturer, this was a fantastic experiment for NTUAI. Admittedly, there were some flaws—for instance, the venue was a bit too cramped, leading to a distracting mix of voices, and we were at a board game cafe but didn’t actually have time to play any games (this is crucial!). However, this opportunity allowed for a real connection between the lecturers and participants. During last Friday’s event, I could truly feel a vibrant and joyful energy in the air. ...

In the world of LLMs, we often hear that “Multi-step Reasoning” or “Multi-agent Architectures” can significantly improve performance. For scenarios involving creativity and brainstorming, allowing Insights Agents to “elaborate more” indeed expands the semantic boundary and sparks unexpected insights. However, in “quantitative finance validation” scenarios, where extreme precision is required, over-elaboration might actually make the performance terrible. I recently read 《A Multi-Agent Framework for Quantitative Finance》, a paper published by JPMorgan at EMNLP 2025. This paper proposes a framework comprising “Insights Agents” such as a Data Summarizer, Finance Expert, and Query Refiner, attempting to enhance the performance of a Base Agent by incorporating financial knowledge and data preprocessing. ...

A year ago today, I was eating a Princi cream bun at Milan Cathedral, admiring the Duomo, and casually scrolling through news about a major power outage in Spain while lying in Sempione Park. That night I was supposed to fly back to Barcelona, and I was wondering if Barcelona Airport would even reopen. Now, this year, I’m reviewing questions submitted by students from the probability TA course I’m assisting with, and rushing to finish my machine learning assignment. However, all machine learning assignments come with a lot of fragmented free time. This model I’m working on takes about 10 minutes to train, and there are many different variants to prepare. After looking at the model, I need to check papers; after checking papers, I need to assess data reasonableness. Even if something seems unreasonable, I’ll discuss it with AI to make it sound plausible. Finally, I also have to keep the report within six pages. I quietly deleted some experimental results that seemed unimportant but were interesting to me, feeling a bit empty. ...

1. ☕️ Let’s Talk AI, Let’s Talk Life: Free Coffee Chat Booking Stop with the superficial self-congratulations; come use a coffee chat to navigate your career uncertainties. I firmly believe that even the most powerful technology remains just cold code if it isn’t translated and discussed. I value a questioning spirit, rigorous technical approaches, and genuine collaboration. If you’re also weary of the common sycophancy in the AI community and yearn for ‘Zero-bullshit,’ direct-to-the-core exchanges, feel free to reach out. ...

Recently, on a whim, I backed up my ChatGPT conversation history, wanting to look back and see where this one-to-two-year journey of “human-computer collaboration” truly began. So, I used ChatGPT’s backup feature. In the backup file, I found what was nominally my “first question”: 時間： 2023-03-23 19:22:14 主題： 翻譯文章 內容： “Can you help me translate several articles from Traditional Chinese to English” It looks perfectly reasonable: a university student, a translation task – a standard starting point for AI use. But looking at this date, something felt off: No, my “first time” was definitely not in March, but earlier, in late 2022. The topic was the “emotional blackmail detector.” ...

If you often discuss decisions with AI, you must have encountered this situation: it always tries to give an answer that “covers all aspects.” Superficially comprehensive, it’s actually a bland, noncommittal, “colorful mud.” This phenomenon stems from AI’s underlying training mechanism (RLHF), which inherently predisposes it to the “mediocrity trap” and “flatterer tendency.” In other words, if your question is mediocre, or lacks specific emotions and strong assumptions, AI will typically tend to give a conservative, comprehensive, but somewhat vague answer. ...