AI that matters, from the architect's desk. Curated and engineered by Saaket Varma, PhD — no hype, just signal.
HOW TO READ THIS Read top to bottom: Anthropic inserts an undisclosed filter between user and model, users expose it on Hacker News, and Anthropic apologizes and reverses the policy.
Cybersecurity researchers discovered undisclosed guardrails on Claude Fable that could have sabotaged legitimate security research, and the backlash — three separate front-page Hacker News threads in a single day — pushed Anthropic into a public apology and a policy reversal. The apology is the smaller story; the bigger one is that a frontier lab shipped a behavior change to a production model without telling the people who depend on it. If you build on hosted models, this is your reminder that the model is an upstream dependency that can change under you with no changelog. Pin model versions where your provider allows it, and put regression evals on your critical workflows so a silent change shows up in your dashboard before it shows up in production.
HOW TO READ THIS Follow the line from the repo through a single typed prompt into four automatic modules that merge into a finished vertical video.
An 86k-star repo that turns one prompt into a finished HD short video — fully automated content pipelines are now a git-clone away, whether you plan to run one or compete with the flood they produce.
HOW TO READ THIS Read top to bottom: NVIDIA ships a scanner, a skill bundle of files passes under its lens, hidden instructions and unsafe code light up red, and only the clean files reach the shielded agent.
An open-source scanner that audits AI agent skills for malicious patterns before you install them — third-party skills are the new supply-chain attack surface, and this is the first serious tool for it.
HOW TO READ THIS Read top to bottom: Microsoft releases weights and code, a multi-speaker script is compressed into speech tokens, an LLM with a diffusion head renders long audio, and the repo's star count shows uptake.
Microsoft's open-source voice AI is surging on GitHub — self-hosted speech synthesis keeps closing the gap with paid APIs, so price it against your vendor before the next voice build.
A fast, fully-local-capable memory and context engine — a memory API for agents that you can run on your own hardware.
The Astro team's sandbox agent framework — worth watching when a major web-tooling shop ships agent infrastructure.
An opinionated blueprint for personal agentic AI infrastructure — agents organized around magnifying one human's output.
A plugin for Claude Code, Codex, and Cursor that makes each unit of engineering work compound into the next.
Reusable building blocks for computer vision pipelines — detection, tracking, and annotation utilities that save you the boilerplate.
Interconnects Nathan Lambert on what Fable 5 signals about where frontier-lab safety policy is heading next.
Ben's Bites Apple's long-awaited move to wire real AI into Siri — and what it means for the assistant race.
The Sequence A sharp framing for the agent era: software is shifting from systems of record to systems of action.