AI / Models / Agents

Learning more about Claude’s mathematical capabilities โ€” Anthropic describes how an unreleased Claude research model raised the verified lower bound on the proportion of Riemann zeta zeros satisfying the hypothesis from 41.6% to 67.2%, with a paper and Lean formalization.

Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models โ€” Dyna Robotics presents a world-action model pretrained on more than one million hours of human video and studies its scaling behavior on human and robot tasks.

[2608.09867] Stealing Reasoning Traces from Proprietary LLM APIs โ€” demonstrates how reusing encrypted reasoning blocks across sessions may expose proprietary reasoning traces, and discusses prompt-injection risks and mitigations.

Introducing Unsloth Desktop โ€” an open-source desktop application for Windows, macOS, and Linux that runs, trains, exports, and deploys several kinds of AI models locally.

MiniMaxAI/MiniMax-Music3 ยท Hugging Face โ€” MiniMax’s model and weights page for its music-generation model.

Training AI Scientists to Replicate Research ยท Inherent โ€” introduces Faraday, a 27B AI research agent trained with long-horizon reinforcement learning to reproduce figures from research papers.

Measuring Autonomous AI Research โ€” reports an evaluation of 18 frontier models across 153 autonomous runs on a nanoGPT optimizer speedrun task.

SlimeBoyOwO/LingChat โ€” an AI-powered galgame chat application with expressive characters, a desktop pet, scheduling, and interactive story modules.

Mathematics / Research / Learning

Problems ยท TheoremDB โ€” a database of mathematics and theoretical computer science problems filterable by field, solution status, and Lean verification status.

[2606.03953] Introduction to stochastic gradient methods โ€” a systematic introduction to deterministic and stochastic first-order optimization, covering smoothness, convexity, the PL condition, convergence rates, and variance reduction.

ComfyResearch: See How Learning Happens โ€” represents data, models, training, and observations as composable graph nodes so researchers can visualize and compare neural-network learning dynamics.

Hadamard Matrix of Order 668 | Epoch AI โ€” a FrontierMath problem on constructing a Hadamard matrix of order 668; the page records a candidate solution developed collaboratively by three researchers and Claude.

Matrix Calculus โ€” an online symbolic tool that differentiates vector, matrix, and tensor expressions and exports results to Python or LaTeX.

GENO โ€” describes optimization problems in a modeling language and automatically generates Python solvers for CPUs or GPUs.

How to keep thinking โ€” considers how frequently reviewing LLM output can affect habits of slow thought, and recommends writing and reading dense books to preserve deeper thinking.

Free eBooks | Project Gutenberg โ€” a free ebook library focused primarily on older works whose U.S. copyrights have expired.

Sakai Association of Science Summer Training 2026 โ€” the 2026 summer training course site for publishing preparatory material, lecture notes, and assignments.

ChapterPal โ€” AI-assisted reading and note-taking โ€” an AI-assisted reading and note-taking tool with contextual Q&A, active recall, and incremental reading.

Math Translations โ€” a community for openly preserving, sharing, and collaboratively maintaining Chinese translations of mathematics texts, lecture notes, and terminology.

Algorithm Wiki โ€” a knowledge base that organizes algorithm material by field, problem, algorithm family, and variant.

Programming / Systems / AI Tools

jcode โ€” an open-source terminal coding agent written in Rust, emphasizing low resource usage, parallel sessions, and an auditable, customizable runtime.

QwenLM/Qwen-MM-Plugins โ€” multimodal plugins for Qwen models and multiple agent frameworks, covering images, video, OCR, search, media editing, CAD, and more.

Optimising fp32 GEMM on A100 โ€” progressively explains and benchmarks Tensor Core, PTX, shared-memory, asynchronous-loading, and pipeline optimizations for fp32 GEMM on NVIDIA A100 GPUs.

lidge-jun/opencodex โ€” a local proxy that translates Codex and Claude Code client protocols to multiple LLM providers, with streaming, tool calls, and account routing.

duolahypercho/codex-router โ€” a local router that connects external Anthropic, Kimi, DeepSeek, and other models to Codex’s native model-selection flow.

PI from Scratch โ€” a Chinese tutorial for building a personal pi-agent from scratch; the page is currently unavailable, so this description relies only on its public title.

lassejlv/gpui-starter โ€” a minimal Rust desktop application template built with Zed’s GPUI, including a basic theme and interaction example.

Beautiful UI โ€” Crafted primitives for AI-native interfaces โ€” copyable interface primitives for AI-native applications, covering streaming states, approvals, tool calls, tasks, diffs, tables, and chat interactions.

raullenchai/Rapid-MLX โ€” a local model-serving engine for Apple Silicon with OpenAI- and Anthropic-compatible APIs for coding agents and other clients.

Waku โ€” one native app for all your coding agents โ€” a native Rust and GPUI application that manages sessions, tool activity, and Git checkpoints for multiple coding agents in one interface.

pytorch-training-recipe ยท ml-training ยท pytorch-lightning โ€” three agent skills for training workflows, covering reproducible PyTorch training, general machine-learning checks, and PyTorch Lightning project organization.

DeepSeek Harness โ€” DeepSeek’s open-source agent runtime with an everything-is-a-plugin architecture, currently in developer preview.

Package Catalog ยท Pi โ€” a catalog of npm-distributed extensions, skills, prompt templates, and themes for the Pi coding agent.

zhu1090093659/dsh-web-ui โ€” a community collection of plugins and skins for the DeepSeek Harness Web UI, including a task board, Git graph, and mobile remote interface.

kenn-io/agentsview โ€” a local-first tool for searching and analyzing coding-agent sessions and token usage.

Zeron โ€” monitors and controls local Claude Code, Codex, Grok, Hermes, and Pi agents from another device in real time.

DannyMac180/grok-plugin โ€” uses MCP and xAI OAuth to let Claude Code or Codex delegate tasks and code reviews to Grok.

xiaobright/dsh-anchored-standard โ€” an experimental community DeepSeek Harness preset that stages the tool catalog to shape initial session behavior before enabling the full toolset.

dshfind โ€” a Chinese community that aggregates DeepSeek Harness plugins and offers plugin metrics, courses, and paper explainers.

blader/humanizer โ€” an agent skill for reducing common signs of AI-generated prose.

Maximizing the value of your Claude Code sessions โ€” Anthropic’s guidance on getting more value from Claude Code sessions, with an emphasis on using tokens effectively.

Home | mise-en-place โ€” a polyglot CLI for managing project tool versions, environment variables, and tasks.

GPUI Component โ€” a component library for building cross-platform desktop applications with Rust and GPUI.

Welcome to Omarchy! โ€” The Omarchy Manual โ€” the official manual for Omarchy, a Linux distribution built on Arch, Hyprland, and Quickshell.

ast-grep โ€” a fast structural code search, linting, and batch-rewrite tool for multiple programming languages.

macOS / Productivity / Other

Tinycast โ€” a lightweight native macOS launcher built with SwiftUI and AppKit, with app launching, hotkeys, clipboard history, and snippets.

History of Market โ€” interactive charts covering the long-term performance, returns, valuations, and drawdowns of the S&P 500, Nasdaq, and other U.S. markets. image.png

VibeCleaner โ€” a native macOS utility that scans and classifies Xcode, package-manager, Docker, Ollama, and other development caches by cleanup safety.

Student Education Discount Directory โ€” a directory of student discounts and education-email verification links for software, shopping, entertainment, and AI tools.

metedata/Remarc โ€” captures feedback on a Mac through text, screenshots, web elements, or voice and passes it to an agent over MCP.

Savant โ€” a note-taking application for Mac and iPhone, currently in early alpha.

Moeka โ€” a Markdown editor focused on a fluid, high-performance editing experience.