AI / Models / Agents

Reward is not the optimization target — LessWrong: Explores the distinction between reinforcement learning rewards and the goals a model learns. In an update, the author acknowledges that reward can empirically become a primary goal, though this does not follow by definition.

AI Alignment Forum: A community for AI alignment research and discussion.

Bayes’ rule — LessWrong: An introduction to Bayes’ rule and the intuition behind it.

LessWrong: Articles and discussions about rationality, cognition, and AI.

Dario Amodei — We Must Pace the Frontier: Dario Amodei discusses coordinating the pace of frontier AI development to give safety research and society more time to respond.

GPT 6 Astra Embodied Policy Evaluation: An anonymous evaluation report on GPT-6 Astra’s embodied policies.

Latest research - Goodfire: Goodfire’s research directory, focused on model interpretability and internal mechanisms.

Augmented Lagrangian Predictive Coding: training 1000-layer networks without backpropagation: Sakana AI applies an augmented Lagrangian method to predictive coding, exploring ways to train deep networks without backpropagation.

P(doom) | Armin Ronacher’s Thoughts and Writings: Armin Ronacher reflects on catastrophic AI risk and how people express its probability.

The Age of Wonders and Terrors | Scott Aaronson: Scott Aaronson discusses the hopes and concerns raised by AI progress.

Introducing System One Models & Jev | TypeSafe AI: TypeSafe AI introduces System One models and Jev for structured decision-making.

From Academia to Alignment | Mike Winer: Mike Winer shares his experience and reflections on moving from academic research to AI alignment.

Existential Risk from AI: An Exposition for Mathematicians: An explanation of AI existential risk and alignment for readers with a mathematical background.

MiMo-v2.6 RL: Xiaomi’s introduction to reinforcement learning for MiMo-v2.6.

Introducing Arrow 2 and Arrow 2 Telos – QuiverAI: QuiverAI introduces the Arrow 2 and Arrow 2 Telos vector graphics generation models.

Training Search Agents with GRPO — Jasper Lu: A practical account of training search agents with GRPO.

Helix 2.5: Zero-Shot 30-Home Generalization: Figure evaluates Helix 2.5 on three household tasks in 30 unseen homes, reporting an overall success rate of 56%; the tasks were fine-tuned in other environments.

Cache-to-Cache: Direct Semantic Communication Between Large Language Models: Research on transferring intermediate semantic representations directly between language models through KV caches, reducing the information loss of text-only communication.

Pretraining data, not verifiability, is why LLMs are especially good at math (and coding) — LessWrong: Proposes that the structure of pretraining data may explain language models’ strengths in mathematics and programming better than the verifiability of answers.

Reading a Pretraining Run — Chunyuan Deng: A guide to monitoring loss, gradients, activations, and evaluations during pretraining, including metric definitions, sampling methods, and diagnostic priorities.

Qwen3.8-livetranslate: The official introduction to Qwen3.8-livetranslate.

Mathematics / Research / Learning

Stanford CS 312 | Deep Learning Alchemy: A Stanford course on practical deep learning training and experimental methods.

Scaling law tutorial: A tutorial on neural network scaling laws.

A beginning for mathematics · Daniel Litt: Daniel Litt discusses the future of mathematical research in the age of AI and the value of understanding, communication, and new questions.

Professor Bingsheng He: Professor Bingsheng He’s academic homepage.

AI in Science | Google: Google’s report on AI’s role in advancing scientific research.

Building Labs that Learn – Periodic Labs: Periodic Labs outlines a research approach in which laboratories keep learning from experimental feedback.

Nature Is Our Learning Environment – Periodic Labs: Discusses using feedback from real-world experiments to train models.

AI Infrastructure at Periodic – Periodic Labs: Introduces the AI infrastructure supporting scientific agents, experiments, and training.

Fast math/slow math | Benjamin Antieau: Benjamin Antieau discusses two rhythms of mathematical research: solving problems quickly and building understanding slowly.

Conjectures.io: Expresses open mathematical conjectures as precise Lean statements and accepts verifiable proofs or counterexamples.

A Short Walk Through the Garden of Mathematics: Matrix67’s popular mathematics book, A Short Walk Through the Garden of Mathematics.

The Fate of the Riemann Hypothesis: Revisited – Proofs and Prompts: Richard Evan Schwartz revisits his short story about the Riemann hypothesis and discusses AI’s effects on mathematical research, education, and careers.

Top 500 Open Problems by LLM-assessed Importance — ProofAtlas: Ranks 500 open problems by importance as assessed by language models; the ranking does not represent a consensus among mathematicians.

Sur la destruction des maths par les boîtes d’IA | David Madore: David Madore comments on the relationship between AI companies and mathematical research.

Solving a Rubik’s Cube with graph theory | The Math Flow: A post explaining Rubik’s Cube solving through graph theory.

An Introduction to Universality | Alessandro Codello: Alessandro Codello’s introductory lectures on universality, covering fractals, percolation, chaos, and phase transitions.

Impossible Questions | Scientific American: A Scientific American collection exploring the origin of life, longevity, dark matter, consciousness, and the beginning of the universe.

I compiled these articles into Chinese–English bilingual PDFs, with separate layouts for phones and computers: Mobile edition (PDF, 103 pages) · Desktop edition (PDF, 56 pages).

  • Finding meaning in the impossible (editor’s introduction): Pursuing seemingly unanswerable questions can lead to unexpected practical discoveries; the journey itself may be meaningful.
  • The questions science may never answer: Science offers approximations that keep evolving. Some questions may remain out of reach because their subjects cannot be observed, or because observation changes them.
  • Will we ever know how life began?: Experiments can reproduce possible paths to life’s origins, such as metabolism-first scenarios or an RNA world, but establishing which path actually occurred is much harder.
  • Will we ever achieve immortality?: Explores slowing aging and cellular reprogramming. Current longevity records do not establish an unbreakable biological limit, though substantially extending life remains a major challenge.
  • What if we never find dark matter?: Gravitational evidence for dark matter is extensive, but its particle identity has not been directly confirmed. Perhaps science can offer predictions without reaching an absolute truth.
  • The hunt for reality’s smallest pieces: Could strings or quantum units of spacetime lie beneath the Standard Model’s 17 particles? Experiments approaching the Planck scale remain almost out of reach.
  • What if ‘consciousness’ isn’t real?: Consciousness may not be a useful scientific category; studying infants, bees, or AI might call for different distinctions.
  • What came before the big bang?: No-boundary models, a big bounce, and a CPT-mirrored universe offer testable predictions, but “before” may be meaningless or impossible to access directly.
  • Why we love unexplainable mysteries: An interview with Unexplainable about the appeal of mysteries, from the universe’s origins to why bicycles stay upright. The unknown need not be discouraging.

Programming / Systems / AI Tools

lulu-loopp/folio-terminal: A terminal for Windows and macOS that brings formula rendering, file previews, and agents’ waiting status into the workspace.

zyycn/codex-proxy-rs: A self-hosted transparent proxy gateway for multiple Codex accounts, written in Rust.

My AGENTS.md: A Life Spent Wrestling with Unnatural Prose and Overengineering: Notes on using AGENTS.md to guide models’ writing and curb overengineering.

Why TeX is Slow and How We Rebuilt It in Pure Rust | Leo Liu · leoliu0/ratex: An introduction to TeX’s performance bottlenecks and RaTeX, a typesetting engine rewritten in Rust.

MiniMax-AI/minimax-code: An open-source terminal coding agent powered by MiniMax.

Ryoku-dev/ryoku-arch: An Arch Linux desktop distribution combining Hyprland, an installer, and system configuration.

The last mile of a long road: faster NumPy in the browser | Notebook.link: The engineering work behind making NumPy run faster in the browser.

UXSpice: A circuit simulator built with Svelte, WebGPU, and ngspice compiled to WebAssembly.

Design / macOS / Other Finds

Qingjian Input Method: Adds foreign-language vocabulary to Chinese typing, helping users pick up words along the way.

Design Studio AI: A design workspace built for agents.

The Component Gallery: Organizes examples from real design systems by component type, making it easier to compare implementations.

Libraries.dev: UI effects libraries that coding agents can install and configure, including glowing borders and liquid effects.

saragordic/window-sweaters: A macOS menu bar app that adds knitted borders to windows.

livabble: A guide to cities and everyday life for Chinese-speaking digital nomads.

KiwiCanopy/KiwiDesk: A macOS tiling window manager with multiple layouts configured through a settings interface, while preserving native desktops.

Savant: The Savant project homepage.

jaywcjlove/key-clicker: Adds mechanical keyboard and typewriter sounds to typing on macOS.

ObsidianUI: A component library built with React and Tailwind CSS.

Compositor: An image editor for Mac.

Design inspiration bookmarks:

  • Mobbin: A large collection of real app interfaces for studying design patterns.
  • Godly: A web design inspiration gallery, now redirecting to Recent.
  • Awwwards: Website awards recognizing outstanding design and execution.
  • Cosmos: Collect and organize visual inspiration, with search by color and visual similarity.
  • Curated Design: A curated collection of websites organized by category.
  • Design Spells: An archive of delightful interface details.
  • 60fps: A gallery focused on interaction and polished animation.
  • Supahero: Website hero sections for quick inspiration.
  • Saaspo: Design inspiration for SaaS websites and landing pages.
  • Layers: A showcase of designers’ work.
  • Minimal Gallery: A handpicked web design gallery.
  • MNMM: An archive of minimalist design.
  • Search System: An archive of graphic and editorial design.
  • Rebrand Gallery: Rebranding projects with before-and-after comparisons.
  • Seesaw: A curated web design gallery updated daily.
  • Same Energy: Image search based on visual similarity.
  • Spiral: A spiral visual feed for discovering inspiration.
  • Game UI Database: A database of game interfaces organized by genre.
  • Interface in Game: Game interfaces organized by element type.
  • HUDs and GUIs: Futuristic interfaces from films and games.
  • Pinterest: A longstanding favorite that remains useful.
  • Loadmo: An archive of creative mobile web design.