AI / Models / Agents

Live Training Run · BrushArena: BrushArena’s live training page.

Introducing Julia 1 | Supersonic Labs: Supersonic Labs’ official introduction to Julia 1.

Introducing Claude Sonnet 5.5 | Anthropic: Anthropic introduces Sonnet 5.5’s improvements in coding, knowledge work, and running costs.

Introducing GPT-6.1 Sol | OpenAI: OpenAI introduces GPT-6.1 Sol and its updates in coding, computer use, and professional tasks.

Exploring Similarity Search with Prompted Subspaces — Jasper Lu: Uses text prompts to construct concept subspaces for image similarity search by color, style, or meaning.

RL is an evolutionary algorithm: Discusses reinforcement learning through the lens of evolutionary algorithms.

Introducing Gemini 4: Google introduces Gemini 4 Argon with an output limit of one million tokens, initially rolling out to trusted cyber defenders and testers.

ExplorationBench: Evaluates autonomous exploration in unfamiliar programming and logic environments, using interpreters and proof checkers to verify results.

Grokipedia: The Grokipedia homepage.

Cogentic: The Cogentic project homepage.

How many AI agents could we run? | Epoch AI: Estimates that high-bandwidth memory shipped in 2025–2027, fully deployed and allocated to the relevant workloads, could support nonstop agents providing weekly working hours equivalent to roughly 140–720 million full-time employees.

AI chips shipped through 2027 could run tens to hundreds of millions of concurrent frontier-model agents. Running nonstop, these agents would supply as many weekly working hours as about 140–720 million full-time employees.

A model guide for the GPT-6 family | OpenAI: A guide to model selection, instructions, context and cost management, long tasks, and production deployment with the GPT-6 family.

Lessons from building an automated research scaffold — LessWrong: Shares lessons from an automated research scaffold: limited productivity gains, alongside failures involving reward gaming, result interpretation, and audit monitoring.

What if automating AI R&D triggers an intelligence explosion? — CASP: Examines whether automated AI research could trigger an intelligence explosion, its implications, and policy responses, while emphasizing uncertainty.

Tokenization: A Survey for Modern NLP | alphaXiv · PDF: A survey of subword tokenization methods, their limitations, and alternatives in modern natural language processing.

Language Model “Shape” | Alex L. Zhang: Explores designing language-model input and output structures around agent scaffolds.

Mathematics / Research / Learning

synthetic-sciences/openscience: An open-source AI workbench for scientific research.

CSAPP Chinese Reading: A Chinese reading site for Computer Systems: A Programmer’s Perspective.

AGMAI: September 29 Statement: The mathematics advisory group draws on community feedback to recommend practices for publishing AI-generated mathematical results and supporting human understanding and review.

Transformer Explainer: An interactive visualization of GPT-2 small that explains computations inside a Transformer.

AI solves a ‘holy grail’ problem from probability theory | Scientific American: Reports on an Anthropic model’s proposed proof of a percolation continuity conjecture and discusses its implications for mathematics.

Apex Leaderboard: The Apex leaderboard announcement page.

MIT 6.5940 Fall 2026 TinyML and Efficient AI Computing: MIT’s fall 2026 course covering pruning, quantization, model serving, distributed training, and deployment on edge devices.

Expenditure Horizon | METR: Measures optimization ability through the intersection of human and agent cost–benefit curves, with an initial study of NanoGPT optimization.

Solving Open Research Problems Together | Meta AI Research: Describes Muse Spark’s collaboration with mathematicians on open research problems, with results checked by independent mathematicians.

Learning Forever using Artificial Neural Networks: Shibhansh Dohare’s doctoral thesis on loss of plasticity, continual backpropagation, and policy collapse.

Programming / Systems / AI Tools

A shared cache for Cargo builds | mr boxington: A shared Cargo build cache that reduces repeated compilation.

deepseek-ai/DeepGEMM-Ascend: A matrix multiplication kernel library for Huawei Ascend processors.

Introducing cf | Cloudflare: Cloudflare introduces an agent-oriented command-line tool for its full API.

Ely · GPUI components: A showcase of GPUI components.

claude.dev: A technical blog for developers building with Claude.

MonoCode: A graphical interface for multiple coding agents.

MyGo: A Go framework for desktop apps with web frontends or native interfaces.

Pi 1.0 | Earendil: Introduces Pi 1.0, an extensible agent framework.

Comma: A continuously running agent for personal tasks.

Lorca: Runs AI teammates locally to work on files and tasks through individual or group chats.

I’m Using GPUI For Everything | Freek Van der Herten: Recommends a Syntax team video about developing with GPUI.

Design / macOS / Other Finds

mg-chao/snow-apps: The source repository for Snow Shot and Snow Image Viewer.

Cida: Translates and polishes text across Mac apps.

tornikegomareli/omarchy-spaces: Shows app icons for each workspace in the Omarchy bar and supports workspace switching.

Pal-AI-Lab/Coopanion: Coo is a desktop companion at the bottom of the screen, supporting chat, voice, and authorized computer use.

The Making of Cursor’s Icons | Minor Adventures: Shares the design process behind Cursor’s icons.

IPPure: Checks IP information, risk, network egress, and leaks.

arXiv: Updated Submission Rate-Limit Policy: arXiv’s official announcement of its updated submission rate-limit policy.

adileo/squirreldisk: A cross-platform disk usage analyzer written in Rust, with sunburst and treemap views.

Yab: yet another browser: The Yab browser homepage.

AI Haotan: Compares prices and stock across marketplaces for AI accounts.

metasequoiaime/msime-windows: The Windows repository for Metasequoia IME, including its input framework, server, interface, and installer.

gillesgoetsch/omacvm: Runs Omarchy in a Mac virtual machine, integrating trackpad gestures, audio, and keyboard behavior.

lxj5820/dsh-boot-animation: Replaces the DSH boot screen with a full-window video that fades into the app.