AI / Models / Agents
Diachronic: Explores human behavior modeling and evaluation with Kairos 1.
I Quit OpenAI Because Its Culture Is Broken | The Atlantic: David Robinson explains why he left OpenAI, criticizes its safety culture, and argues for learning from other industries that manage serious risks.
Introducing Beam | Reflection: Reflection introduces Beam, a 501-billion-parameter mixture-of-experts model for coding, reasoning, and agent tasks, with weights planned for release later this month.
Rho-1: Collapsing the multimodal stack | Reka: Reka previews Rho-1, which combines text, images, video, and actions in a single model.
Explainability as a form of system closure: Frames explanation as building approximately closed models over variables humans can grasp, using chaotic systems to explore the limits of interpretability.
Introducing Reason Agent 1.0 路 Reason Agent: Reason Machines introduces Reason Agent 1.0 for autonomous software engineering.
Introducing TasteVal | P-Zero Research 路 TasteVal: Measuring the Experimental Research Taste of AI Systems Against Human Experts: Evaluates how models design experiments and interpret results on fixed research problems, comparing experimental compute efficiency with human experts.
Mathematics / Research / Learning
Practical Parametric Yield | Irrational Analysis: Explains chip parametric yield through process variation, clock speed, and power, then discusses the tradeoffs facing Cerebras’s wafer-scale computing.
A Riemann Hypothesis Explainer Made with Opus 5.5: The uploader presents a Riemann hypothesis explainer with video made using Claude Opus 5.5 and narration provided by Gemini.
Two Room-Temperature Antiferromagnetic Semiconductor Candidates | Vals AI: Reports two candidate room-temperature antiferromagnetic semiconductors found with Claude agents, sharing calculations and limitations; their properties still need experimental validation.
Presenting a proof of the Kannan-Lovasz-Simonovits conjecture: Bizeul, Klartag, and Lehec present a 32-page manuscript proposing a complete proof of the KLS conjecture, disclosing that ChatGPT supplied most of the proof and mathematical ideas.
An $O(1)$ Bound for the KLS Constant: Song and Zhang propose an $O(1)$ upper bound for the KLS constant in the second version of their preprint, claiming to resolve the full conjecture.
A Dimension-Free Bound on the Poincar茅 Constant of Isotropic Log-Concave Measures: Balasubramanian and Kasiviswanathan release another 53-page manuscript claiming to prove the KLS conjecture, also disclosing AI involvement.
Stark units for real quadratic fields and reciprocity laws: Proposes reciprocity laws for Stark units in real quadratic fields, claiming to answer this case of Hilbert’s twelfth problem and derive Zauner’s existence conjecture in every dimension.
The Squares Project 路 n = 11 Case: Collects constructions, lower bounds, and machine verification records for packing eleven unit squares into a square; the project reports matching bounds, with its review record still pending.
The Fibonacci Geometry of Eleven Squares (PDF): Describes Trump’s eleven-square packing using flat tori and Fibonacci geometry, analyzing the unique minimum within its two-parameter contact family.
Sharing AI progress in mathematics | OpenAI 路 openai/math: OpenAI releases manuscripts, proofs, and verification materials from AI-assisted mathematical research, noting that the results still require review by the mathematical community.
Beyond n log n 路 Integer multiplication: Tracks conditional improvements to integer multiplication complexity bounds and related verification materials.
CrocSwap/integer-mult-bounds: Investigates conditional improvements to the exponent saving in OpenAI’s integer multiplication result #109.
The Mathocalypse | Scott Aaronson: Scott Aaronson discusses recent AI mathematics results and their implications for research, verification, and mathematicians.
What does an improved upper bound on $\omega$ mean | Abbas Mehrabian: Explains what improving the upper bound on the matrix multiplication exponent $\omega$ means.
On Matrix Multiplication Algorithms 路 Fast Matrix Multiplication (PDF): A lecture and survey on matrix multiplication algorithms, covering fast algorithms and complexity research.
Erich Friedman 路 Erich’s Packing Center 路 Cubes in Cubes: Erich Friedman’s mathematical puzzles and geometric packing resources, including records for packing cubes into a cube.
Nopert Polyhedra Explorer: Interactively explores polyhedra in Rupert’s problem through rotation, cutting holes, and passage tests.
Relativity Park: Walking Speed of Light: Sets the speed of light to walking speed so visitors can observe relativistic effects in a park scene.
Soft-to-Rigid Packing: Compares packing searches that gradually harden soft shapes with other methods, sharing experiment logs, code, replays, and their respective strengths and limitations.
News | Yasuyuki Kawahigashi: Yasuyuki Kawahigashi records operator algebra researchers’ reactions to OpenAI manuscripts and progress toward proofs humans can read.
A Brief Digest of Some New OpenAI Results | Persiflage: Frank Calegari comments on some newly released OpenAI mathematics results.
Big O Speedrun: Tracks complexity bounds for integer multiplication, Fourier transforms, 3SUM, all-pairs shortest paths, and other problems, distinguishing claims from verification status.
A reaction to the reactions | Bal谩zs Pozsgay: Responds to discussions about AI mathematics from a theoretical physicist’s perspective and reflects on becoming a student again.
Programming / Systems / AI Tools
egoist/godiff: A native Git diff viewer for reviewing working trees, commits, and GitHub pull requests.
sahenjp/rustdsh: A Rust launcher for DeepSeek Harness with configuration migration and local tools.
How fast is Python 3.15? | Miguel Grinberg: Compares the Python 3.15 release candidate with older interpreters using recursive and sorting programs, including just-in-time compilation and free-threading modes.
sshx: Shares terminals live in the browser for collaboration.
Public testing is beginning | Superlogical: Rex begins inviting public testers, initially on macOS, with persistent sessions and remote connections.
egoist/gorex: A Rex terminal clone built with Go and MyGo’s native interface.
ArtCraft Crafting Apps 路 GitHub 路 VectorCraft 路 WordCraft 路 DeckCraft 路 CADCraft: A collection of open-source creative applications built in Rust for vector graphics, documents, presentations, and computer-aided design.
palemoky/chinese-poetry-api: A project providing Chinese poetry search and an API service.
Freaction/Aquilum: A local Markdown knowledge base for Windows with live preview, search, a relationship graph, a reader, and Model Context Protocol support.
solidsf-inc/sfFFT: An MIT-licensed fused Tensor Core fast Fourier transform convolution kernel for NVIDIA GB10.
Ely Office 路 ZacharyZhang-NY/ElyOffice: The Ely Office website and its Rust / GPUI desktop office suite project link; the repository currently returns 404.
RelayKit | New API: A Go module that converts requests, responses, and streaming events among four text protocols from OpenAI, Anthropic, and Gemini.
Tern | Stencil: A native Rust terminal and multiplexer compatible with Kitty, offering persistent sessions and cross-device connections; currently in closed testing.
There are many themes, but this one is yours | Earendil: Explains how Pi reads terminal colors to generate its system theme, discussing palettes and interface contrast.
Rewriting Prime Agent in Rust: Prime Intellect shares how multiple agents rewrote Prime Agent from TypeScript into Rust and how the team verified the rewrite.
Tov: An experimental language for coding agents with TypeScript-style syntax, native binary compilation, and structured compiler errors with suggested fixes.
Design / macOS / Other Finds
RevPDF: A PDF tool for offline editing, comparison, and format conversion.
An Opus 5.5 Video Recreation | R: Shares an Opus 5.5 creation in response to a video recreation challenge.
Paper Mono: A free, open-source monospace font with variable weights, coding ligatures, and a glyph showcase.
celve/uvim: A menu-bar application that brings Vim’s modal editing to macOS text fields.
AIME: An open-source Chinese input method for macOS built on RIME, with dictionary updates, saved phrases, and translation and rewriting of selected text.
Designeer: A curated directory of interface design tools, components, and inspiration.
alejandrobujan/tendedero: A native macOS app that hangs screenshots on a line along the top of the screen for temporary reference.
Pulse: Displays processes, processor usage, memory, network, energy use, and other system metrics in one native window.
Nongfsq/ftop: A floating system monitor for Apple Silicon Macs showing processor, memory, network, and process information.