Date: Jan 25, 2026, Author: Zhaocheng Zhu (co-written with Gemini)
What does it mean to be an AI researcher? If I were asked when I first started my AI journey, I probably would have described it as a career that chases the latest technology, invents cool models that automate the world, and earns wide recognition with a decent salary. Yet, after eleven years, I’ve realized that vision is far from enough to describe the complex path that I have traveled.
Lately, many junior students have reached out to me with questions that sound familiar: “What should I expect for my PhD years?” “How can I make things work and get paper accepted?” “Shall I pursue a career in academia or industry?” When I look into their eyes, I see my younger self—full of ambition and energy, yet unclear about the situations ahead and the moves they can take.
This blog is not a “how-to” guide for publishing at top-tier conferences or landing a research job. Instead, it is a collection of my experience and reflections on questions that I struggled with during each period. It’s about papers that never got published, nights I spent doubting myself, and the moments I had to choose between following the crowd or following my heart. I wrote this for the version of me that existed eleven years ago, hoping that these insights might help you navigate your own path.
It’s a long read taking about 19 minutes, so I’ve added some of my photos along the way to make it feel more vivid and engaging.
When I began my undergraduate research in 2015, deep learning was rapidly transforming the field, and many classical AI benchmarks were being largely boosted by CNNs and RNNs. It was really surprising that one can let machines learn decisions from raw data, while textbooks just taught us to program machines to work. Soon I found playing with these techniques far more interesting than doing my coursework. I spent months hacking into the C/C++ implementations of word2vec, trying to figure out what machines had learnt in those magical representations.
One of my first aha moments came during an experiment with RNNs: I discovered that using Pinyin (the romanization system for Chinese) worked better than word segmentation for tokenizing Chinese input. It was a cool discovery driven by pure curiosity. However, at that time, I had no idea what a research cycle should look like. Consequently, that project, along with some others, ended up nowhere, leaving my CV with several project descriptions but only one experimental arXiv paper.
The turning point came during a summer internship at Mitsubishi in Japan. For the first time, I was out of the classroom and inside a real research lab. I observed two different working styles among researchers. Some were particularly strong at brainstorming and convincing others with their ideas, while others excelled at execution and implementation. I found myself drawn more towards the former, mostly because I was only comfortable at carrying out ideas when I knew why they would work. It seemed like that I probably need a PhD degree to be one in the former group.
By the time I prepared the materials for PhD programs, I faced a cold reality: I had zero accepted papers. While many of my peers had a second-author or a co-first-author paper, I only had many folders of experimental code, along with a head of unproductive curiosity and unrealized taste. It felt like I had wasted my most critical years.