A short list of articles still worth opening — not every post in the archive.
AI & Technology
Ox Alpha is a newly released anonymous AI model with a 1M-token context window, multimodal input, and coding-agent capabilities. Here is what makes it special—and the leading theories about who built it.

AI Research
Agent Lightning v1.0 brings harnessed agentic RL to real coding agents—~3,500 lines of code, near-zero rewrites, and a 14.6-point SWE-bench Verified jump for Qwen3.5-9B with only ~6K examples.

AI & Technology
DeepSeek Harness v0.1 is an MIT-licensed, local-first open-source agent runtime built on Cordis—where models, tools, sessions, sandboxes, and the loop itself are all swappable plugins.

AI Security
Research shows attackers can extract hidden reasoning traces from proprietary LLM APIs across Anthropic, OpenAI, and Google—posing intellectual property and privacy risks.

One note per piece, so the page reads like an edit instead of a grid of products.
AI Industry
Meta unveils Muse Glimmer—a 30-billion-parameter Apache 2.0 open-weight model optimized for always-on local agent workflows on consumer hardware, with up to 20,000 tokens/sec on a single NVIDIA GPU.
AI Infrastructure
Cloudflare unveils Kitesurf, the world's first agent-first browser—running on Workers with V8 isolates, using 3–7× less CPU and memory than Chromium for AI agent tasks.
AI & Technology
Compare Crawl4AI, Firecrawl, Scrapy, and Apify for AI-ready web crawling in 2026—speed, features, pricing, and which tool fits RAG, enterprise scraping, or managed pipelines.
AI Infrastructure
How Cloudflare's breakthrough optimization techniques are transforming large language model deployment—41% higher throughput and 30% lower cost for Kimi and GLM on Workers AI.
AI & Technology
QM YC (yc-software/qm) surged past 10.9k GitHub stars in about six days—and advances beyond OpenClaw and Hermes with multi-agent orchestration, hierarchical memory, adaptive tooling, and enterprise-grade planning.
AI & Technology
Berkeley Sky Computing Lab’s open-source PixelRAG retrieves over 30 million web pages as screenshot tiles—preserving tables, charts, and layout that text RAG destroys—with 97% storage savings and a live Wikipedia endpoint.