AI / Models / Agents
moonshotai/Kimi-K3 · Hugging Face — Moonshot’s Kimi K3 model weights and model card.
ali on X: “22580: From GPT2 to Kimi3, Explained” — a walkthrough of architecture evolution from GPT-2 to Kimi K3.
cneuralnetwork/smol-kimi-k3 — a ~49M-parameter Kimi-K3-style small model trainable on a single 8GB GPU.
cneuralnetwork/smol-glm-5.2 — a ~50M-parameter GLM-5.2 architecture study with MLA, DSA, IndexShare, sparse MoE, shared MTP, and Muon Split.
Exclusive | Nvidia Invests in Sutskever’s Safe Superintelligence AI Lab - WSJ — Nvidia invests in Ilya Sutskever’s SSI to expand compute reach and collaboration.
Compression and Intelligence — Ryan Greene — on the relationship among compression, intelligence, and LLMs.
Physics of Language Models — Allen Zhu et al.’s “physics of language models” series, explaining LLM behavior via architecture and training mechanisms.
Deconstructing Scaling Laws: Optimization, Architecture, and Data - Scientific Spaces — unpacks scaling laws through optimization, architecture, and data.
MinghuiChen43/awesome-deep-phenomena — a curated, continually updated list of interesting empirical deep-learning phenomena and papers.
AI Explorables | PAIR — interactive AI concept explorables from Google PAIR that visualize model mechanisms.
Looped Transformer | alphaXiv — a paper collection on looped Transformers.
Towards Looped Models Done Right — on designing and training looped / recurrent-style models properly.
Explorative Modeling — Unlocking a Third Pretraining Axis and End-to-End Generation | Alexi Gladstone — proposes explorative modeling as a third pretraining axis. Core idea sketch:
losses = []
for i in range(K):
generation = model.generate() # e.g., from a different random noise
losses.append(loss_fn(generation, data))
min(losses).backward() # only the best generation gets gradients
What optimization dynamics appear if $\min_{i} \mathcal{L}_{i}$ is generalized to an arbitrary $f(\mathcal{L}_{1},\dots,\mathcal{L}_{K})$?
[2604.22951] The Power of Power Law: Asymmetry Enables Compositional Reasoning — argues that the power-law asymmetry of natural language helps compositional reasoning and should not simply be flattened.
Enigma: Robots that feel right — Enigma on robots that understand their surroundings and act in ways that “feel right.”
maka-agent/maka-agent — a local-first AI desktop assistant.
ArXivIQ | Grigory Sapunov | Substack — hand-picked ML papers with AI-generated deep dives; daily TLDR and architecture notes.
Blog - Ulam · Research - Ulam — AI research notes and projects such as DeepAlgebra.
Home | CSE 234 — course page for CSE 234 Machine Learning Systems.
Mathematics / Research / Learning
Mathematics in the age of AI — Terence Tao’s slides on mathematics in the age of AI (ICM 2026 context).
What Is the Inverse Covariance Matrix? — explains the precision (inverse covariance) matrix from statistical and geometric viewpoints.
[2607.21716] A Defense of the Quadratic Model — defends the quadratic model in optimization theory as a tradeoff between tractability and real loss landscapes.
AI Agents for the Working Mathematician — practical AI-agent workflows for working mathematicians.
Prompting Toward a Conjecture - Guanyang Wang — how one extra sentence about searching the literature changed GPT-5.6 Sol Ultra’s success rate on Feige’s $1/e$ conjecture.
Generating Proofs with LLMs and Completing Formal Verification - Guanyang Wang — LLM-generated proofs plus Lean formal verification, illustrated on the binary-matrix case of the 1997 KTV conjecture.
Ten advances in mathematics and theoretical computer science | OpenAI — OpenAI’s summary of ten advances in mathematics and theoretical computer science.
UnsolvedMath — a repository of open mathematics problems.
Home | Open Problem Garden — a directory and community garden of open problems in mathematics.
Optimization Online — an eprint repository for optimization and related topics.
Trouble in High-Dimensional Land - Dibya’s Blog — counterintuitive phenomena in high-dimensional geometry and probability.
faridrashidi/cnsplots — a toolkit for Cell / Nature / Science-style publication-quality plots.
Programming / Systems / AI Tools
Herdr: one terminal for the whole herd — an agent multiplexer: one terminal for a whole herd of agents.
Shuffle - A fast, native file manager for macOS — a free open-source native macOS file manager (Rust + GPU rendering) with dual panes, tabs, command palette, and terminal mode.
Tinycast — a tiny, native macOS launcher — a lightweight native macOS launcher with fuzzy search, calculator, clipboard history, and global hotkeys.
Rust Vs C++ - LLM coded model conversion from PyTorch — LLM-written PyTorch model-conversion logic compared side by side in Rust and C++.
icebird1998/drawio-scientific-illustrator — step-by-step scientific illustration via MCP control of a live draw.io canvas.
zeronsh/chat — a unified AI chat app spanning Claude, OpenAI, Gemini, and more.
terminal-browser — a real browser that runs inside the terminal.
Codex Reset — Last Reset Tracker & Live Forecast — tracks Codex usage-reset times with 24h / 48h forecasts.
GPU Glossary — Modal’s glossary of GPU terms.
ascend-ai-coding/awesome-ascend-skills — a Huawei Ascend NPU knowledge base organized as distributed Agent Skills.
Doubao Search Custom — Volcengine Doubao Search docs; 500 free calls per month.
iv-org/invidious — an alternative YouTube front end.
Design / Writing / Other
://argon.red — a designer’s blog; related discussion on X from Alpha Argon.