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: Amazon and Anthropic start linked, Jassy carries the issue to Washington, Washington's ban grounds Anthropic, and the partner bond ends cracked.
The Wall Street Journal reports that Andy Jassy's talks with U.S. officials — citing Amazon's own cybersecurity research — drove the White House order forcing Anthropic to cut global access to Fable 5 and Mythos 5. That reframes the export crackdown from a Washington-vs-AI-labs story into a partner-vs-partner one: Amazon is Anthropic's single largest backer, yet pushed the case that its models were too dangerous to ship abroad. The signal for anyone building on frontier models is that your most strategic investor and your regulator can be the same pressure, and model availability is now a geopolitical variable, not a roadmap line. Treat single-vendor frontier access as a supply-chain risk: know which capabilities you depend on, and have a fallback before access decisions are made above your vendor's head.
HOW TO READ THIS Read top to bottom: Bezos funds Prometheus, the startup swaps chat replies for an agent that plans, codes, runs and retries until working software ships.
Bezos' new startup Prometheus reportedly targets an 'artificial general engineer' — confirming the frontier race is shifting from chatbots to agents that actually build.
HOW TO READ THIS Read top to bottom: a folder of files feeds graphify, which links the files into a node graph, driving 67,451 stars for the still-research project.
A skill for Claude Code, Cursor, Codex, and Gemini CLI that turns any folder of code, docs, or schemas into a queryable knowledge graph — a concrete fix for the 'my agent has no memory of my codebase' problem.
HOW TO READ THIS Read top to bottom: Apple ships Foundation Models, and processing now runs inside the device's own boundary instead of reaching the crossed-out cloud, powering Siri and Foundation Models apps.
Apple's Foundation Models and an upgraded Siri are rolling out on-device — the biggest real-world test yet of edge AI running in consumers' pockets.
Open-source deep-learning framework for building and fine-tuning physics-ML models — NVIDIA's bet on AI for simulation and the physical world.
Open infrastructure for computer-use agents: sandboxes, SDKs, and benchmarks to train and evaluate AI that operates real software.
Fully local long-term memory for AI agents via a 4-tier progressive pipeline — the other half of the agent-memory problem graphify is chasing.
Extracted system prompts from Claude, ChatGPT, and others — a useful primary-source reference for how frontier assistants are actually steered.
A maintained list of free LLM inference endpoints accessible via API — handy when you need a fallback after a model access decision goes sideways.