AI / Models / Agents
Former OpenAI reasoning lead JT and Gemini pretraining lead Rohan on the future of continual learning — a conversation about continual learning, post-pretraining knowledge updates, and long-term model adaptability.
(Some of) The Models, They Just Don’t Want to Learn | Tilde — examines potential recursive depth and training difficulties in reasoning models, with the One Layer Deeper competition exploring the joint design of architectures, optimizers, and objectives.
Maple-Preview Model Card — DeepGrove — DeepGrove’s open-source 20B-A1B reasoning model with ternary weights, designed for fast on-device inference and adaptation.
Xiaomi-Robotics-1 — Xiaomi’s robot foundation model, pretrained on more than 100,000 hours of embodiment-agnostic manipulation trajectories, then aligned with target embodiments and instructions using real-robot data.
Lattice: an 8 MB static retriever that embeds Wikipedia in 7 minutes — an approximately 8 MB static retrieval model and pure-Rust runtime that can embed 6.4 million English Wikipedia articles in about seven minutes on an 8-core M2 MacBook Air.
[2606.12683] From AGI to ASI — outlines four possible paths from AGI to ASI: scaling, paradigm shifts, recursive improvement, and large-scale multi-agent collectives.
What are AI robots? A complete introductory guide to 2026 — an introductory guide to AI robots, their core components, common types, and applications.
Evolution Strategies at the Hyperscale — introduces EGGROLL, which uses structured low-rank perturbations to increase the GPU arithmetic intensity of evolution strategies; the authors report roughly 100× higher throughput than naive ES at billion-parameter scale and large population sizes, with experiments on pure-integer RNN pretraining and RWKV reasoning post-training.
Mathematics / Research / Learning
The Daily Diff — a daily collection of worthwhile papers, technical articles, and discussions with recommendation signals and concise summaries.
TensorTonic — an interactive platform for learning machine learning by implementing algorithms, CUDA kernels, and landmark papers from scratch.
Interactive Machine Learning — Prof Richard Xu — an online course that explains machine-learning concepts and algorithms through interactive material.
Inference in Statistical Modelling and Machine Learning — the companion site for a textbook on inference in statistical modeling and machine learning, with a complete PDF, solutions, slides, and Python / R notebooks.
Inference Engineering — an interactive guide to AI inference spanning GPUs, CUDA kernels, quantization, KV caches, parallelism, and production deployment.
Kurate.org — AI Paper Rankings — a paper-ranking site where three LLMs jointly assess the potential impact of preprints in robotics, economics, physics, and other fields.
AI Settles a 25 Year-old Problem We Left Behind — discusses the binary integer least-squares problem underlying maximum-likelihood MIMO detection and an example of AI-assisted progress on related theory. Exact solution in general remains NP-hard; polynomial-time results usually refer to relaxations or approximate detectors. The target problem is:
$$ \min_{\boldsymbol{x}\in\{\pm 1\}^{N}} \lVert \boldsymbol{y}-\boldsymbol{H}\boldsymbol{x} \rVert_{2}^{2} $$JoseDeFreitas/awesome-youtubers — a curated list of technical YouTube channels covering web development, computer science, machine learning, game development, cybersecurity, and more.
Programming / Systems / AI Tools
jcode — an open-source terminal coding agent written in Rust, emphasizing low resource usage, parallel sessions, persistent long-running work, and auditability.
VisualTeX — an online tool for visually editing and previewing LaTeX formulas in real time.
Peezy — a terminal coding agent for real projects with a unified interface for switching among multiple models and OpenAI-compatible providers.
Diffs, from Pierre — an open-source code and diff rendering library built on Shiki, with split and unified views, annotations, conflict handling, and interactive accept-or-reject controls.
yetidevworks/drydock — a Rust TUI for checking uncommitted changes, unpushed commits, repositories behind their remotes, and unreleased versions across multiple repositories.
Canario — a macOS terminal built on the Rio engine, with spatial grouping, splits, a command palette, tab previews, and a global quick terminal.
Local Studio — a local AI workspace that brings models, inference runtimes, hardware monitoring, files, terminals, and agents into one interface, with unified access to vLLM, SGLang, MLX, and llama.cpp.
CNB — Tencent’s cloud-native development platform, managing build environments, caches, plugins, and elastic compute through an Everything as Code model.
SteveSuv/remix-words-funny — a full-stack English vocabulary-learning project built with React Router v7, tRPC, Drizzle, PostgreSQL, HeroUI, and Docker.
codecast-sh/codecast — records, searches, and remotely steers Claude Code, Codex, Cursor, and Gemini sessions, with team memory and line-level agent attribution.
macOS / Productivity / Other
Tahul/space-rabbit — removes macOS Spaces switching animations to reduce desktop-transition delays.
Vorssaint · vorssaint/vorssaint-utils — a free, open-source modular macOS menu-bar toolkit for sleep prevention, hardware monitoring, volume, networking, disks, and common system controls.
Side Screen — turns an Android tablet into an extended display for a Mac over USB-C or Wi-Fi, with hardware acceleration and touch input.
migsilva89/imark — a native macOS Markdown reader that refreshes as files are saved; annotations are stored directly as HTML comments in the Markdown file without uploading it to the cloud.