First-Year Master's Recap

In November 2025, I had never seriously looked at a single line of code. My daily computer use was opening WPS; when it came to anything computer-related, I was basically starting from zero.

Ten months later, I’m preparing my thesis proposal.

I’m impatient, I’m anxious, I want too much. I want to grow faster, faster, and I feel like I could go even faster. This post isn’t a success story written for other people; it’s an inventory I’m doing for myself — laying out what actually happened each month, including the time I wasted.

The first half year: from 0 to “I guess I’ve studied everything”

November 2025: cramming the fundamentals in one go

The first month was the most intense. I spent over two weeks on Mu Li’s Dive into Deep Learning, and what stuck with me most was implementing CNN, RNN and Transformer from scratch — not calling APIs, but writing the architectures myself. At the same time I went through FunRec’s recommendation fundamentals and the machine learning basics in the “Watermelon Book,” and I also read Hands-on Reinforcement Learning, going through the basic RL algorithms once.

December: the full recommender system pipeline

Before December 20, I studied FunRec’s news recommendation project and got the whole pipeline working end to end: retrieval, ranking, re-ranking.

Then I read countless interview-prep cheat sheets from senior students, all kinds of foundational knowledge points. Now? I’ve basically forgotten all of it.

But I don’t think it was wasted. Having seen something and never having seen it are two different things — things you forget because you don’t use them often come back quickly; things that never entered your head at all are always brand new the first time.

January 26 to March 4: wasted

Winter break. I barely studied. Roughly five weeks, just gone.

March: running into vibe coding

After getting back to campus, this was my first encounter with vibe coding. Claude Code, Codex, and Antigravity suddenly flooded into my awareness. I researched every peripheral thing there was: how to use relay services, how to farm Plus accounts, tools for managing Codex accounts, how to hook up the Claude Code desktop version to an API, how to run Codex on a server.

March also had some legitimate work:

• Built my own project on the KuaiRand dataset: cross-domain ranking plus multi-interest fusion. (Honestly, I bought this project, lol. I was too impatient, desperate to land an internship.)
• Got MiniOneRec running and understood the overall logic of generative recommendation

April: a month of distraction

Got into Obsidian, kept tinkering with my toolchain. Went off and studied Java for a few days, didn’t stick with it. Then Datawhale’s Agent & LLM course, plus assignment one of CS336.

May: exam month

Busy with all kinds of exams, busy with my Party membership defense, became a Party development candidate. Basically didn’t study before May 20.

Started on the TAAC competition on May 22, and picked LeetCode back up at the same time — after two months away, I’d forgotten most of it.

Half-year summary

I really did make progress over these six months, going from zero to being able to run a complete recommender system pipeline. But the problem is obvious: my goals weren’t focused enough, and my hunger wasn’t strong enough.

I spread myself too thin — deep learning, recommendation, reinforcement learning, Agents, Java, terminal tools, a bit of everything. At the time I listed three abilities I genuinely needed to build up:

1. Communication ability
2. Papers
3. Coding ability

Not one of those three was seriously trained in those six months.

June: turning what I’d learned into something solid

This month the focus shifted from “learning new things” to “turning what I’d learned into something I can show”:

• Systematically learned Claude Code fundamentals
• Built my own personal blog
• Wrote project summaries, wrote my resume, prepared for interviews
• Hello Agent, plus ACM-style whiteboard coding practice

On June 15, I became a probationary Party member.

July: half competitions, half the world getting bigger

I read through CS336 assignment two once; honestly I didn’t understand it and just moved on. I signed up for AFAC and the Kuaishou LLMRec competition, but didn’t see either through.

From July 6 to August 28, I interned at PRT, mainly working on dataset construction and designing appropriate rubrics.

July 19: my first social at Shanghai Tower

August

Starting July 19, I took on the benchmark and self-evolving router project Max gave me. I worked on it for a while, then he left, and I stopped too.

A pile of scattered, useless experience: using Agents to delete memory, organizing browser bookmarks, API format protocols, the Ghostty / Yazi / tmux terminal toolset, my first encounter with WAIC and embodied intelligence, using TransOne to translate and read English papers.

Personal life: on August 12 I got together with Zhuzhu. On August 21 I got in touch with Professor Ma at Renmin University about a research internship. On September 6 I met Zhuzhu in person for the first time — she bravely came to find me at Jinhua Station when I was heading back to school.

August 29 - September 8: preparing the thesis proposal

Which is right now.

Reflections

In ten months I got through a lot of courses, a lot of books, a lot of interview cheat sheets, but there isn’t much I can actually show someone: one KuaiRand project, one working MiniOneRec, one blog, one internship. I signed up for three competitions and didn’t see a single one through.

Compounding, keep at it every day

Do less fake work: researching relay services, account management tools, terminal configuration — these things give instant feedback, a sense of control, and they look very “technical.” But they don’t produce any result that can be verified. I fell into this trap in both March and April.

Get out and socialize more: that social on July 19 — the people I met and the possibilities I glimpsed in one day were worth more than a month of studying with my head down. Part of the anxiety comes from having too narrow a view — if you don’t know which direction to push in, all you can do is push in every direction.

Underneath it all, there isn’t a seed of my own, something I’d want to chase even with no one alongside me. A person with no internal question is destined to rely on external people to define “what to do right now” — and when those people aren’t around, they just spin in place.

Every direction I took this summer died from “switching to yet another one,” but they all died for the same reason: there was no seed of my own underneath. It broke off the moment the UCB friend left, not because the project was bad, but because it wasn’t my question.

“Find a question I’d want to know the answer to even if no one were watching. Is there something that, even with no one pushing me, no one looking, no deadline, I’d still sit down and start working on the moment I thought of it?” Not for interviews, not for a resume, purely being hooked by something and wanting to know “what’s actually going on here”? Maybe that’s what grad school is: a lot of questions stop having standard answers, and you have to go find the answers yourself.

Someone with industry value, resources, and an understanding of the market won’t be easily replaced — they have know-how of their own. From technical to management/TL, socializing, seeing things from other people’s perspective Subtract; you need product sense, business sense, communication, taste, AI workflow ability, SOPs What the top people are doing, what VCs are investing in My problems: anxiety from comparing myself to others, setting too many goals at once, not sleeping properly

With my shallow learning and narrow vision, it’s hard to see the full picture of things objectively

Watch out for

Eyes, head, skin, sleep, teeth, diet, reading, exercise, side projects