Xiaoyu Li

Mission log · Earth · T − 00:00:10

Xiaoyu Li

What does it mean for AI to help a person?

I’m trying to understand myself and the world I find myself in. Mathematics, art, and philosophy are ways I explore it.PhD student · UNSW Computer Science and Engineering

Launch

Scroll to fly. Fast scrolling is warp speed; move the pointer to steer; click to send a probe.

Low Earth orbit · 400 km up

Mission objectives

Three questions to carry the whole way.

  1. OBJ-1 · Purpose

    What should AI help people pursue?

    How can people express, revise, and refuse goals? Learning a preference is not permission to act.

    Goals · Revision · Authorization

  2. OBJ-2 · Agency

    When does help empower rather than replace?

    Does assistance increase a person’s ability to understand and act? People should be able to choose how much to delegate and take back control.

    Assistance · Capability · Control

  3. OBJ-3 · Evidence

    How do we know the help was real?

    What observations justify a claim of benefit? When do guarantees survive connections between models, tools, and people?

    Identification · Verification · Composition

Full mission brief

Saturn · about 1.4 billion km from home

Transmissions

Messages from home, received near the rings. Each one is an essay.

  1. RX 025The New Eight-Legged Essays and the Smiling, Proud Wanderer新八股与笑傲江湖
  2. RX 024Why Is the Golden Snitch Worth 150 Points?金色飞贼为什么值150分？
  3. RX 023Can't Learn It? Is It the Concept, or the Teaching?学不会，是概念太难，还是讲法错了？
  4. RX 022Werewolf with a Million Players: 674 Wolves Are Enough一百万人的狼人杀，674 只狼就够了
  5. RX 021Why Don’t We Design Slower Algorithms?
  6. RX 020You Didn't Learn Badly. You Were Taught Too Well.你不是没学好，是被教得太好了
  7. RX 019Why You Always Lose at Tetris俄罗斯方块，为什么注定会输
  8. RX 018Chess on an Infinite Board无穷棋盘上的国际象棋
  9. RX 017Above Infinity无穷之上
  10. RX 016A Question with No Answer一个答不出的问题
  11. RX 015Counting Past ω数到 ω 之后
  12. RX 014I See It, but I Don't Believe It我看见了，但我不信
  13. RX 013Infinities Come in Sizes无穷也分大小
  14. RX 012The Universe Did Not Include a Progress Bar
  15. RX 011Isn’t It Just Math?不就是数学嘛？十个箭头，和箭头背后的东西
  16. RX 010722 Math Papers Land on GitHub Overnight722 篇数学论文，一夜上了 GitHub
  17. RX 009AI Can Prove Theorems Now. Why Are We Still Writing Math?AI 都会证明了，我们还写数学干嘛？
  18. RX 008Is English the Best Programming Language?英语是最好的编程语言吗？画家、钟表匠和巫师
  19. RX 007Why I Asked AI to Build a PEFT Atlas
  20. RX 006What Changes When AI Helps Us Write?
  21. RX 005Does Dota 2 Armor Really Diminish?DOTA2 的护甲，到底递不递减？
  22. RX 004In 48,000 Tang Poems, What Did Poets Write About Most?四万八千首唐诗里，诗人最常写的是什么？
  23. RX 003From the Four Color Theorem to Navier–Stokes从四色定理到 NS 方程：机器证明的五十年
  24. RX 002Reason and Feeling: Really Opposites?理性与感性，真的是对立面吗？
  25. RX 001Hofstadter’s ABCD: Between Creation and Discovery侯世达的 ABCD：在创造与发现之间
The whole archive of transmissions

Beyond the wormhole · from here the map is fiction

The planets of a black hole

Six worlds in orbit. The nearer you go, the faster time runs at home.

  1. Planet 1 · Technical report · 2026

    A Categorical Atlas of Parameter-Efficient Fine-Tuning: Cones, Orbits, and Gauges

    Hundreds of fine-tuning methods, viewed through one mathematical framework.

    PDFInteractive companion

  2. Planet 2 · arXiv:2606.14688 · 2026

    Flood and Harvest: The Provable Necessity of Trivia for Generating Valuable Mathematics via the Lens of Language Generation in the Limit

    Discovering unseen valuable mathematics requires infinitely many correct but worthless statements.

  3. Planet 3 · arXiv:2607.13003 · 2026

    Watermark Forensics for Generative Models: An Information-Theoretic Perspective

    We quantify the cost of user attribution, payload recovery, and watermark localization.

  4. Planet 4 · arXiv:2608.26187 · 2026

    When the Canonical Completion Is Wrong: Formalizing and Measuring the Jump in LLMs

    We formalize and test whether LLMs can leave a canonical default when constraints demand a different structure.

  5. Planet 5 · arXiv:2609.06862 · 2026

    Feature Superposition in Neural Networks: From Theory to Practice

    We review how neural networks encode more features than dimensions and how to recover them.

  6. Planet 6 · arXiv:2609.38446 · 2026

    What Pretraining and Midtraining Make Learnable from Rewards?

    Task-independent prediction can teach computations that rewards alone leave undetermined.

Selected papersThe 2026 archive

Inside the tesseract

Works

A room built of every moment of a bookshelf. Reach through.

Every work

Home · the pale blue dot

Send a message home

The beacon is saying hello in Morse code. Answer it.

Other ways in: the arcade · the book · the game world · the deep sea · the theorem · plain text

Fly again

Distances: Earth orbit ≈ 400 km, the Moon ≈ 384,400 km, Saturn ≈ 1.2–1.7 billion km. The wormhole, the black hole and the tesseract are borrowed from fiction. Time dilation for an observer held still near a black hole: Δthome = Δt / √(1 − rs/r).