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: the lab, the checkpoint it built, the download, and who can now run it.
Thinking Machines released Inkling, an open-weights 975B-parameter LLM — its first public model after a year and a half of building infrastructure out of view, and it immediately topped Hacker News. A frontier-scale checkpoint you can download changes the calculus: the open-vs-closed debate stops being theoretical when the weights are sitting on your disk. For teams locked into closed APIs, this is the strongest self-hosted bargaining chip yet — and for regulated environments, a serious candidate for workloads that can't leave the boundary. Pull the weights, benchmark it against your current production model on your own evals, and price out what serving it would actually cost before your next contract renewal.
HOW TO READ THIS Read top to bottom: a stolen credential breaches Suno's system, revealing that YouTube videos were scraped and fed into Suno's model, confirming the scraping claim.
Data exposed via a stolen employee credential shows Suno trained on millions of songs scraped from YouTube, Genius, and Deezer. Training-data lawsuits have mostly run on inference and expert testimony — this is rare hard evidence, and it lands mid-litigation. If you're building on generative audio or video APIs, provenance just became a procurement question, not a philosophical one: ask your vendors for training-data attestations in writing, because indemnification clauses are about to get tested.
HOW TO READ THIS Read top to bottom: OpenAI acts, ships the $230 Codex keyboard, does so while the Apple lawsuit stays active, and lands its first hardware product.
Mid legal fight with Apple over hardware trade secrets, OpenAI shipped its first device: a $230 light-up keyboard purpose-built for its Codex coding agent, with a screenless ChatGPT speaker reportedly coming this year. The bet is that agents earn dedicated hardware, not just app real estate — a physical invoke button is a distribution play aimed straight at developer muscle memory. The takeaway isn't the keyboard; it's that OpenAI thinks the agent interface war will be fought on your desk, and lock-in now has a hardware layer.
HOW TO READ THIS Read top to bottom: the maker ships the repo, its pipeline turns raw input into finished video, then the star count shows its traction.
A fully automated open-source AI short-video engine — topic in, finished clip out — currently surging on GitHub trending at 25,579 stars. The end-to-end content-factory stack that startups charge SaaS pricing for is now free to self-host, which compresses the moat for every AI video tool overnight. If you ship video at any scale, spend an hour with the pipeline architecture even if you never deploy it — the orchestration patterns are the interesting part.
Fullstack MCP framework for building MCP apps for ChatGPT/Claude and MCP servers for agents.
Agentic RL training at scale — infrastructure for training agents, not just prompting them.
Stripe's official toolkit for building AI-powered products with payments wired in from day one.
Personal trading agent from HKU's data science lab — autonomous market analysis in agent form.
iMessage personal agent running on Claude Agent SDK or Codex runtime, with persistent memory.
The Sequence OpenAI's own analysis of where coding evals break — worth reading before you trust a leaderboard delta.
Ben's Bites Practical guide to GPT-5.6 — useful baseline for comparing against whatever you're running today.