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 ships Opus 5, it sits close to Fable 5 but cheaper, landing it #1 on Artificial Analysis.
Anthropic released Claude Opus 5 — billed as cheaper and less restrictive than Fable, with capabilities 'close' to Fable 5 — and it promptly took #1 on the Artificial Analysis Intelligence Leaderboard. The combination matters more than the ranking: frontier-class output at a lower price resets the default model choice for anyone building on Claude, whether direct API or Bedrock. Re-run your evals against Opus 5 this week; if it clears your quality bar, reserve the premium tier for the narrow set of workloads that actually need it.
HOW TO READ THIS Read top to bottom: OpenAI and Anthropic push Washington for a crackdown, the industry forks into two camps, then Nvidia, Microsoft and Meta join the opposing side.
Startup founders urged Washington not to shut off Chinese open-weight models, backed by a Nvidia/Microsoft/Meta letter warning against overregulation — while OpenAI and Anthropic lined up on the other side. The alignment is the story: the labs selling closed frontier models want restrictions; the companies selling chips and infrastructure don't. If your roadmap assumes continued access to open weights, treat that as a policy risk now, not a given — model availability is a supply-chain dependency like any other.
HOW TO READ THIS Read top to bottom: bfl.ai's FLUX 3 core, then it out-guns Seedance 2.0 and Gemini Omni, then FLUX-mimic turns its video frames into motion that drives a robot, ending in robotics now being a shipped target.
Black Forest Labs launched FLUX 3, multimodal flow models reported to beat Seedance 2.0, Gemini Omni and Grok Imagine — plus FLUX-mimic, a video-action model aimed at robotics. A challenger lab out-benchmarking the biggest names in generative video is notable on its own; pointing the same stack at robotics is the bigger signal. Video-action models are shaping up as the bridge between generative AI and physical systems — worth watching even if embodied AI isn't on your roadmap yet.
HOW TO READ THIS Read top to bottom: a plain box becomes a server after one install, gains a full stack, and earns 3,609 stars.
ODS is trending on GitHub (3.6K stars): one open-source install that turns a PC, Mac, or Linux box into a full AI server — LLM inference, chat UI, voice, agents, workflows, RAG, and image generation. That's the entire self-hosted stack in one repo, private, on hardware you already own. If you work in a regulated or privacy-sensitive environment, this is the fastest path yet to prototyping an assistant stack that never leaves your network.
Kernel library for LLM serving — the performance layer under your inference stack.
A DSL for writing high-performance GPU/CPU/accelerator kernels without hand-rolling CUDA.
Route, manage, and analyze LLM requests across multiple providers behind one unified API.
A persistent memory layer built for AI agents.
Notion's official MCP server — wire your workspace into any MCP-capable agent.