This page demonstrates every content type the site supports. The source is a plain markdown file (data/pages/en/features.md) — open it in the editor to compare.

Text & typography

Bold, italic, strikethrough, inline code, titled links, and rich interactive footnotes¹. In modern academic and editorial publishing, footnotes provide seamless bibliographic citations² as well as detailed technical annotations³.

In magazine typography, blockquotes are the breathing room. — a typography enthusiast

• Unordered item A

• Unordered item B

  • Nested item

• Done: static site pipeline

• Done: markdown rendering

• Doing: write more content

Code highlight

import torch

def cosine_lr(step: int, total: int, base: float = 3e-4) -> float:
    """Cosine-annealed learning rate."""
    t = min(step / total, 1.0)
    return base * 0.5 * (1 + torch.cos(torch.tensor(t * 3.14159)))

Math

Inline $e^{i\pi} + 1 = 0$, and block-level softmax:

$$\mathrm{softmax}(z_i) = \frac{\exp(z_i)}{\sum\nolimits_{j=1}^{K} \exp(z_j)}$$

Figures & grids

Two-column grid (collapses to one column on mobile):

Embedded players

Players render the official iframe directly in a responsive 16:9 container (loading="lazy", so first paint stays fast):

Self-hosted media (native tags):

Functional directives

GitHub Repo Card:

Streaming Block:

Editorial components

The complete kit below is embedded by ::editorial{id="features"} from editorial_blocks in site.yaml. It covers actions, numbered list cards, tiles, archive cards, and a divider:

03 · Components

EDITORIAL KIT

This block is embedded with ::editorial{id="features"}; it shows actions, list cards, tiles, archive cards, and a divider.

Primary actionSecondary actionQuiet action

02List card with imageMask · HoverHover raises mask opacity and scales the image down slightly.

03List card without imageText · StableThe numbered two-column layout remains stable without an image.

WideWide tileTallTall tileSmallStandard tileSmallStandard tile B

ArchivedArchive card AGrayscale media regains some color on hover.

LegacyArchive card BUseful for retired projects and old experiments.

---

The contact card, QR modal, light/dark theme, language switcher, background music, image lightbox, and scroll reveal are global components; interact with this page to see them.

The homepage-specific Profile Block, GitHub Block, and RSS Block are also rendered in full below; they do not depend on the homepage layout.

Profile Block

Zhiyuan Lin

PhD candidate / ML & Systems

• Email
• GitHub
• Google Scholar

GitHub Block

2,750 GitHub contributions in the last year

SepOctNovDecJanFebMarAprMayJunJulAugSep

MonWedFri

LessMore

stlin256/OpenHomepage-V2Lightweight, magazine-style static personal homepage generatorastrobilingualgithub-pagesmarkdownpersonal-homepageportfolioTypeScript100Updated 1 hour ago

huggingface/transformersState-of-the-art NLP for everyone — infrastructure for reproducing papersaudiodeep-learningdeepseekgemmaglmhacktoberfestPython164.7k34.4kUpdated 3 minutes ago

ggml-org/llama.cppRunning LLMs on a laptop — where local experiments beginggmlC++126.9k22.7kUpdated 6 minutes ago

openai/whisperRobust Speech Recognition via Large-Scale Weak SupervisionPython108.3k13.1kUpdated 12 minutes ago

RSS Block

Frontier Radar

OpenAI

Daybreak for Frontline Defenders: $1B to protect essential servicesOpenAI2026-09-03OpenAI introduces Daybreak for Frontline Defenders. A $1 billion commitment expands access to frontier cyber AI, traininLegora reviewed 41 documents in minutes with GPT-6 AstraOpenAI2026-09-03Legora used GPT-6 Astra to review 41 documents in minutes, find all four planted errors, and improve performance by nearPlayco cut manual fixes 50% prototyping games with GPT-6 AstraOpenAI2026-09-03Using GPT-6 Astra, Playco built three themed game prototypes from one grey box foundation and reported 50% fewer manualSafety overview: GPT-6 AstraOpenAI2026-09-03GPT-6 Astra is our most capable broadly deployed model and our first to reach the Critical level of cybersecurity capabiATV Big Air Tour turned 3 days of work into 3 hours with ChatGPTOpenAI2026-09-02ATV Big Air Tour uses ChatGPT Work to speed up marketing, merchandising, and more. It even turned merchandise photos int

Google DeepMind

Introducing WeatherNext 3, our most advanced and accurate global weather AI modelGoogle DeepMind2026-09-03Proactive cyber defense for governments and enterprisesGoogle DeepMind2026-09-02Introducing Gemini 3.8 Flash and 3.8 Flash CyberGoogle DeepMind2026-09-02Introducing agentic video understanding with GeminiGoogle DeepMind2026-09-01Gemini Omni 1.1 Flash lets you build with more controlGoogle DeepMind2026-08-27

Hugging Face

NeoMME: an efficient Multimodal-native and Multilingual EncoderHugging Face2026-09-03Fine-tuning a 350M Model for Better Structured Outputs in 100 GRPO StepsHugging Face2026-09-03Give Your Coding Agents a Memory You OwnHugging Face2026-09-03

量子位

卡车自动驾驶里程超10亿公里，嬴彻科技定义“货运物理AI”量子位2026-09-04吃下90%市场新版GPT Image 2.5已经能伪造GPT-6发布会了量子位2026-09-03改进GPT Image 2的噪点问题刚刚，GPT-6正式发布！OpenAI：欢迎来到AGI时代量子位2026-09-03全球最强C

Editor's Picks

Co-Scientist: A multi-agent AI partner to accelerate researchEditor's Picks2026-05-12Introducing Co-Scientist, a collaborative AI partner built with Gemini to help researchers accelerate scientific breakthMulti-agent systems accelerating research — directly related to my workFrom games to biology and beyond: 10 years of AlphaGo’s impactEditor's Picks2026-03-09Ten years since AlphaGo, we explore how it is catalyzing scientific discovery and paving a path to AGI.Ten years of AlphaGo — a piece of tech history worth readingGemma Scope 2: helping the AI safety community deepen understanding of complex language model behaviorEditor's Picks2025-12-16Open interpretability tools for language models are now available across the entire Gemma 3 family with the release of GA benchmark for interpretability tooling; this card intentionally has no cover, demoing text-only cards

Page controls

Page controls are page-level widgets configured on a per-page basis (non-global). Each page can define them independently in its frontmatter; they reappear whenever the page is reopened or revisited, providing page-specific announcements and controls.

💡 Example: The homepage of this site features a prominent yellow Notice Banner (notice: { text: "This is a demo page. Content is for displaying project features only.", color: "yellow" }) that appears 0.5s after load.

Raw HTML mixing

This line uses the native HTML mark tag. Dangerous tags like <script> are filtered by a whitelist.

Callouts & timeline

The P0 content directives remain script-free at runtime: callouts explain and warn, while timelines present education, experience, and milestones.

Education & Experience

1. 2022–2026

   PhD Candidate

   Example University

   Focused on machine learning and systems, especially inference scheduling and reproducible evaluation.

2. 2026 · Present

   Research Intern

   Example Lab

   Worked on on-device LLM inference experiments.

Publications & Citations

data/publications.yaml is the canonical source, while publications.bib supplies raw BibTeX by key. Filtering, sorting, and grouping happen at build time; copying BibTeX is the only progressive enhancement.

2026

01

OSDI 2026oral

Efficient Inference with Adaptive Scheduling

Zhiyuan Lin, Alice Doe, Bob Smith

First author; led scheduler design and evaluation.

Abstract

An adaptive inference scheduler for interactive workloads, evaluated with distribution reports on real traces.

@article{lin2026efficient,
  author = {Lin, Zhiyuan and Doe, Alice and Smith, Bob}

02

MLSys Evaluation Workshop

Reproducible Evaluation for Open Models

Alice Doe, Zhiyuan Lin

Co-author; built the evaluation set and reporting scripts.

@inproceedings{doe2026reproducible,
  author = {Doe, Alice and Lin, Zhiyuan}

2025

01

Journal of Machine Learning Systemsartifact

Edge KV Cache Compression

Bob Smith, Zhiyuan Lin

@article{smith2025edge,
  author = {Smith, Bob and Lin, Zhiyuan}

This site natively supports scholarly bibliographic citations and rich interactive footnotes¹. On desktop devices, hover or focus over references to trigger smart popovers containing bibliographic metadata, inline code, and mathematical formulas³; on mobile devices, a smooth bottom drawer sheet slides up, followed by a dedicated footnotes appendix at the end of the document².

Footnotes

1. Vaswani et al. (2017). Attention Is All You Need. Advances in Neural Information Processing Systems (NeurIPS 2017). arXiv:1706.03762 ²

2. Standard CommonMark & GFM footnote extension with desktop popovers and mobile bottom drawers. ²

3. Embed inline code O(N \log N) and math equations like $L = -\sum y \log \hat{y}$ right inside footnotes. ²