About

The name

Chinese names usually come in three characters and are chosen for both sound and meaning. Mine, in order, are:

宋

Sòng — the family name. An ancient surname, also the name of a thousand-year-old dynasty (960–1279) known for its painters, poets, and exam-driven bureaucrats. I inherited the surname, not the dynasty.

仲

Zhòng— "second" or "middle." In the classical sibling-ordering system (伯·仲·叔·季 — eldest, second, third, youngest), 仲 is the second-born. In its later, looser sense it just means in-between, balanced, neither extreme.

轩

Xuān — a high carriage, a balcony with a view, openness and lift. In classical Chinese, scholars often named their study room with the character — 某某轩 — to signal a place of clear thinking with a wide outlook. It is the character that does the aspirational work.

So 仲轩 reads roughly as balanced & with-a-view — a hope more than a description. Johnnycame later, picked because it's easier to call across a noisy room and easier to put on an English mailing list.

Where I'm from

Born in Henan, raised mostly in Hainan after my family moved south.

From 2018 to 2024 I attended Hainan Overseas Chinese Middle School (海南华侨中学) — locally just called 侨中, founded in 1938 — for both junior and senior high. Six years at one school is a long time; a lot of who I am now formed there.

Where I am now

Since August 2024 I've been at the University of Science and Technology of China (USTC)in Hefei, Anhui — a third part of China, neither plains nor coast. I'm currently in the Huaxia Talent Class, an honors track I joined in my second year.

Two years in, two years out — the slight unease and slight optimism that comes with standing in the middle of something.

Three schools in three years

USTC lets undergraduates transfer between schools, and I've used that freedom every year so far:

• Year 1 · 2024–2025 — School of Information
• Year 2 · 2025–2026 — School of Computer Science · joined the Huaxia Talent Class (current)
• Year 3 · 2026–2027 — School of Big Data and Artificial Intelligence (upcoming)

Each move was voluntary. The principle is simple — I take what I'm interested in seriously, and I'd rather change departments to keep chasing a question than stay put for inertia.

What I work on

I've been a research intern in three labs so far. They look different from the outside but share one question: how do information systems hold and retrieve what matters?

1. fMRI brain-computer interfaces — research intern at the USTC Medical Imaging Center, advised by Prof. Xiaoxiao Wang. We decoded image semantics from brain activity. This was where I first saw how a real biological system compresses, routes, and reconstructs signal under heavy noise.
2. Test-time context construction — research intern at MIRA Lab, advised by Prof. Jie Wang. We formalized context selection as a cost-regularized information bottleneck (IBC²). The question shifts from what model should we use to what subset of the world should we put in front of it right now.
3. Path-aware auditable memory for agents — research intern at AlphaLab, advised by Prof. Xiang Wang. SHELF is our attempt to turn agent memory from opaque similarity retrieval into something verifiable: state frames, evidence ledgers, an auditor-gater protocol. You should be able to ask an agent why did you do thatand get a real answer.

The arc, looking back: brain → context → memory. Each internship taught the next one what to ask.

What I build

Outside of papers I ship things — Chrome extensions, desktop tools, voice interfaces, small experiments. A constraint I keep: each project should be something I'd actually use, or something a friend asked for. That filter throws out a lot of ideas, and that's the point.

The four I currently maintain — Anymark (a bookmark agent), PromptFlow (a local-first prompt manager), Burner Note (zero-knowledge self-destructing notes), and smart-voice-chat (a low-latency voice loop) — live on theProjects page.

What I write

I write a Chinese-language WeChat newsletter under the pen name JohnnyXuann. The mantra: “Notes I think matter, before consensus catches up.”

So far it has been: a 60K-character source-code analysis of Claude Code, a cross-strait pricing dissection of Kimi K2 (with a 78-day correction), a read on OpenAI's compute reallocation around Sora, and a few quieter essays on attention and identity in the digital age. Eight original posts at the time of writing.

English posts, when they come, will live on theWriting page.

How I work

I am a heavy user of AI tools — across Claude Code, Codex, Gemini CLI and assorted models, my cumulative usage as of mid-2026 is around $15,800, 171K messages, and 28.4 billion tokens. I record this not as a brag but as a data point: this is the environment I think and build inside, and a lot of my research instincts come from being inside it.

A version of those numbers lives on theToken Stats page.