AI / Models / Agents
The bitter lesson for computer vision: Vincent Sitzmann discusses how computer vision could become part of end-to-end perception and action loops.
Hadamard MLP: A channel-mixing design based on the Hadamard transform explores ways to reduce feed-forward layer parameters.
Qwen-Image-2.1: The official introduction page for Qwen-Image-2.1.
Tokenizers v1 benchmarks: Benchmarks for the Tokenizers v1 release candidate cover encoding, decoding, and parallel performance, with instructions for reproduction.
Grok 4.7: The official announcement describes Grok 4.7 updates for coding and knowledge work.
KDA Doesn’t Care About the Future: An investigation of future-dependent rounding in KDA kernels and whether models exploit that numerical effect.
Halo: A distributed training framework for Hugging Face models with checkpointing and multi-turn reinforcement learning.
Limite 1B — Violetto: Paradigma introduces a one-billion-parameter model for mathematical reasoning and describes its training and inference implementation.
MiMo-V2.6 · MiMo-V2.6 Technical Report: The official introduction to Xiaomi’s MiMo-V2.6 model series.
HySparse2: A hybrid sparse attention architecture with two-level key-value cache sharing for long-context and multi-turn agent tasks.
Jesse: An open-beta service for solidSF’s inference engine with an OpenAI-compatible API.
Batch Size Scaling — Tencent Hunyuan RL Team: The Tencent Hunyuan team discusses how batch size relates to training outcomes when scaling reinforcement learning for language models.
Introducing Claude Opus 5.5 | Anthropic: Anthropic introduces updates to Opus 5.5 in coding, knowledge work, communication, and operating cost.
Introducing GPT-6 Sol and Luna | OpenAI: OpenAI introduces GPT-6 Sol and Luna, covering improvements to capabilities, inference efficiency, and caching.
Qwen Intelligence: The Qwen Intelligence project website.
The Generalization Problem | Q: Explores combining model search with generalization priors so that learning can scale with compute.
Introducing HLE-Diamond: A cleaned and refined set of 1,000 questions from Humanity’s Last Exam, covering reasoning and expert knowledge.
Computational depth is all you need | Q: Studies the relationship between computational depth, model expressiveness, and training efficiency, with experiments on deeper language models.
Visualizing modular-addition training | Sauers · The models just want to learn | Sauers: Compares output distributions after 2,000 and 20,000 training steps, using circular plots to visualize learning in a modular-addition transformer. A second post shares the author’s thoughts on model learning.
Intelligence Density | Trajectory: Discusses intelligence density as task performance per unit of inference compute.
What my deep model doesn’t know | Yarin Gal · Dropout as a Bayesian Approximation: Yarin Gal introduces uncertainty estimation with dropout, alongside the paper interpreting dropout as approximate Bayesian inference.
Mathematics / Research / Learning
The Khachiyan constant and GPT-6 Astra: The author reports on using GPT-6 Astra to explore a conjecture about the Khachiyan constant.
Hilbert space: treating functions as vectors: An introduction to Hilbert spaces starting from the idea of functions as infinite-dimensional vectors.
Memristive singular value decomposition: The authors report singular value decomposition on memristive compute-in-memory chips and discuss applications including image reconstruction and model-weight decomposition.
Hamiltonian Neural Network with PyTorch: A tutorial on training Hamiltonian neural networks in PyTorch through an oscillator example.
Chinese Character Map, second edition · Chinese Character Map: Matrix67’s introduction to the second edition of the Chinese Character Map.
Optimization Constants in Mathematics: A collection of mathematical optimization constants and known upper and lower bounds, with links to research sources.
CIS 6270: Discrete Generative Models: A University of Pennsylvania course covering discrete generative models, diffusion, flow matching, and optimal transport.
MS&E 319: Efficient Generative Language Models: A Stanford course on training objectives, architectures, and inference algorithms under limited computational budgets.
Physics in the age of LLMs: Oz Amram shares his experience and views on how large language models are changing his physics research workflow.
Advisory Group on Mathematics and AI · AGMAI: OpenAI introduces an independent mathematics advisory group and collaboration on reviewing results, communicating research, and developing research tools.
A Proof of the Most Informative Boolean Function Conjecture: The authors propose a computer-assisted proof of the Courtade–Kumar conjecture, with computational verification records and a description of formal verification.
Dictators are most informative: The authors propose a proof of the Courtade–Kumar conjecture concerning the information retained by dictator functions under binary noise.
Unsolved problems in mathematics: A Wikipedia index of open problems by mathematical field, with references for further reading.
devol.dev: Interactive tools for learning robot kinematics, control, motion planning, and design.
A Faster Shortest Path Algorithm | Vals AI: The author reports using Claude agents to develop the C-HD shortest-path algorithm and a Lean proof, improving an asymptotic bound for a specific graph-density range without establishing a practical speedup.
Quanta Magazine: A science magazine covering research in mathematics, physics, biology, and computer science.
Claude discovers a novel enzyme system | Anthropic · Research preprint (PDF): Anthropic reports that Claude identified an enzyme system with CRISPR-like repeats in sequence databases and describes experimental work; its function remains unknown.
Slava Krushkal’s good-group discovery | Logical Intelligence: Logical Intelligence reports on Slava Krushkal and the team’s AI-assisted study of a new good group in four-dimensional manifold topology.
mo271/Zeta5: A Lean 4 formalization project for the irrationality of $\zeta(5)$; the repository reports a completed proof and explains its differences from the preprint.
Interactive Linear Algebra — UBC edition: An interactive linear algebra textbook covering vectors, matrices, linear transformations, eigenvalues, and related topics.
Headlines and inside stories | Tapio Schneider: In a guest post on Terence Tao’s blog, Tapio Schneider discusses turning AI-generated results into understandable, auditable, and trustworthy mathematical and scientific knowledge.
We’re gonna need a lot more mathematicians | Amit Sahai: In a guest post on Terence Tao’s blog, Amit Sahai argues for expanding the mathematical research community to understand AI-generated ideas and preserve human agency.
Claude computes a nine-loop amplitude | Anthropic: Anthropic describes Claude’s calculation of a nine-loop six-particle amplitude in planar $N = 4$ super-Yang–Mills theory and how the result was checked.
A question: does the sequence $n + \tan(n),\ n \in \mathbb{N}$ have a lower bound?
Programming / Systems / AI Tools
GPUI Markdown Studio: A WebAssembly-based Markdown editor and preview with Mermaid diagram support.
LayerFS: LayerFS gives agents isolated disposable filesystem branches with checkpoints, rewinding, and state reuse.
Cloudflare Quick Tunnels: Run cloudflared tunnel --url http://localhost:8000 to create a temporary public URL for a local service and share previews or run tests.
Eli Bendersky’s website: Eli Bendersky’s technical blog covers programming languages, systems, and mathematics.
Reader: An open-source macOS reading workspace for PDF, EPUB, and TXT, with annotations, notes, and AI chat.
Tylina: A Typst editor for editing the typeset page directly, available on the web and desktop with AI agent integration.
The senior engineer death spiral: An essay on overwork driven by the urge to prove oneself, and on maintaining communication and delivery as a senior engineer.
CUDA Matmul Optimization Worklog: A step-by-step CUDA matrix multiplication worklog covering coalesced memory access, shared memory, and tiling.
Hallucination Translation: A Mac document translation app with side-by-side reading that preserves formulas, citations, and layout.
OpenCode Reloaded: An explanation of how OpenCode hot-reloads configuration, tools, and plugins while sessions keep running.
Step Code: A terminal coding agent with code editing, tests, plugins, multi-agent coordination, and static-site publishing.
Nowdex: Tracks remaining usage for several AI tools across Mac, iPhone, and iPad, with a menu bar app, widgets, and iCloud sync.
mujocolab/mjlab: Provides an Isaac Lab-style interface powered by MuJoCo-Warp for reinforcement learning and robotics research.
iAmCorey/Wake: An agent session manager built with Rust and GPUI for browsing, searching, and resuming sessions.
heypcb · Circuit World: Generates schematics, board layouts, and fabrication files from natural-language descriptions, with KiCad projects you can keep. Circuit World showcases user-published boards and open-source hardware.
ReShiki: A tool for drawing molecules, building reaction schemes, and preparing chemical figures.
magpie: Selects models for agents such as Codex and Claude Code through a menu bar app and a local gateway.
Synara: An introduction page for an AI coding workspace for Claude Code, Codex, and Cursor.
Empryo: Maintains a live map of a code repository and reads it before edits, with desktop and terminal interfaces.
o8vm/octox: A Unix-like operating system written in Rust and inspired by xv6-riscv.
AGIdock: A cloud platform for agents that provides infrastructure for running applications.
yu1745/wetype-ime-linux: Ports WeType’s Android input engine to Linux and integrates it with Fcitx 5.
Lorca: Runs AI teammates on your computer to work on files and tasks through individual or group chats.
Using Claude Code: Spending your effort | Thariq: Uses evaluations and practical tasks to examine how reasoning effort affects speed, independent judgment, and verification of edge cases.
kaolti/phantomat: An infinite-canvas window manager for Hyprland with zoomable window layouts.
Design / macOS / Other Finds
Working Title: Working Title is an app for Mac.
Crisp: An open-source macOS display utility for scaling, brightness, volume, and display presets.
21st: A collection of React components, page templates, and themes with previews and implementation code.
HowToLiveBetter: A collection of health, everyday-life, and public-service resources with costs, benefits, and references.
Powermove: A video editor that connects to coding agents and supports layers, keyframes, and custom extensions.
Inspora: A curated archive of recent visual design and creative work.
shawn.: Design engineer Shawn’s portfolio features clean interfaces and interaction details.
Opensource UI: A copy-and-paste React and Next.js component library with interface controls and device mockups.
quicker.chat: A Mac inbox that brings direct messages together and offers conversation organization and reply drafts.
MiMo Gallery: A MiMo showcase presented as a three-dimensional portfolio museum.
Top-Conf Figure Gallery: A gallery of paper teaser figures that can be browsed by conference, year, and visual pattern.
Pelican on a Bike: A real-time 3D demo of a pelican cycling along the coast, with waves, a day–night cycle, and music.
Variora: Compares interactive creations from different AI models given the same brief.
Designeer: A curated collection of interface design tools, component libraries, design systems, and inspiration.
TinyWeb: A small native Mac browser with a borderless interface, vertical tabs, and searchable history.
Search | Office Commun · driceroland/Search: A lightweight macOS browser built on WebKit, with a minimal interface and a focus on quick startup.
Lean Browser: A minimal macOS browser built with Swift and WebKit, offering Zen mode and stating that it collects no telemetry.
Macfolio: Curated software, hardware, workspace setups, and reading for the Mac ecosystem.
Sidekick: Adds a second dock to the Mac for apps, links, clipboard history, and live widgets.
saragordic/rooms: Saves each project’s window layout as a room and switches between workspaces with a shortcut.
PriceRadar: Compares download, in-app purchase, and subscription prices across App Store and Google Play regions, with price alerts.