Skip to the contents

Free play · 8 games

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 in the School of Computer Science and Engineering at the University of New South Wales

Every coloured tile is a game, and most of them hide a page of this site. Pick one up and play.

Wordworm

after Snake

Eat the letters in order and the worm writes out one of my essay titles. Its world is a Möbius strip: go out on the right and you come back on the left, upside down.

Flipwell

after Tetris

Every four lines earn a flip: the whole well turns upside down, every block tumbles to the floor, the holes vanish and full rows cascade away. Each line adds a word to my question.

Hunch to Atlas

after 2048

Merge hunches into notes, lemmas, proofs and papers, all the way to an atlas. Reviewer 2 drops in now and then, and won't merge with anything.

Tiny Planet

after the dinosaur game

Run round a planet small enough to lap. Whatever you jump over comes round again, a little bigger each time, until it blossoms. Catch the paper planes.

Catnap

after Minesweeper

Minesweeper with ten sleeping cats for mines. Lift a cushion with a cat under it and it wakes up; leave a fish to keep one napping. Some cushions hide letters. Wake three and nap time is over.

Lantern

after Breakout

Breakout in the dark: the ball is the only light, and every row is one of my titles. Break a whole title to free a firefly; catch it, and the lantern shines further.

Research

  1. PurposeWhat should AI help people pursue?How can people express, revise, and refuse goals? Learning a preference is not permission to act.
  2. AgencyWhen 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.
  3. EvidenceHow do we know the help was real?What observations justify a claim of benefit? When do guarantees survive connections between models, tools, and people?
The research statement

Desk Hop

after Jump Jump

Hop across a mathematician's desk. A run of dead-centre landings plays the folk song Mo Li Hua, beat by beat; land on a book or a curio to pick up one of my pages.

Ink Crane

after Flappy Bird

Fly a crane through bamboo in an ink painting. Your path dries into calligraphy, your best flight flies beside you as a ghost, and lanterns carry my essays.

Works

Every work

Essays

  1. The New Eight-Legged Essays and the Smiling, Proud Wanderer新八股与笑傲江湖
  2. Why Is the Golden Snitch Worth 150 Points?金色飞贼为什么值150分?
  3. Can'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. Werewolf with a Million Players: 674 Wolves Are Enough一百万人的狼人杀,674 只狼就够了
  5. Why Don’t We Design Slower Algorithms?
  6. You Didn't Learn Badly. You Were Taught Too Well.你不是没学好,是被教得太好了
All 25 essays

Selected papers

  1. Technical report · 2026A Categorical Atlas of Parameter-Efficient Fine-Tuning: Cones, Orbits, and GaugesPDFInteractive companionVideo
  2. arXiv:2606.14688 · 2026Flood and Harvest: The Provable Necessity of Trivia for Generating Valuable Mathematics via the Lens of Language Generation in the Limit
  3. arXiv:2607.13003 · 2026Watermark Forensics for Generative Models: An Information-Theoretic Perspective
  4. arXiv:2608.26187 · 2026When the Canonical Completion Is Wrong: Formalizing and Measuring the Jump in LLMs
  5. arXiv:2609.06862 · 2026Feature Superposition in Neural Networks: From Theory to Practice
  6. arXiv:2609.38446 · 2026What Pretraining and Midtraining Make Learnable from Rewards?

Selected papersThe 2026 archive

Secret levels

Earlier versions of this homepage, still open for anyone who finds them.

Your pocket

Pages you win in the games land here, and stay in this browser.

Say hello