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: Washington reaches into Anthropic's live, actively-serving models, then severs every connection — the first such intervention into a running commercial AI system.
A U.S. export-control directive ordered Anthropic to cut off foreign access to Claude Fable 5 and Mythos 5 — and because it can't filter foreign users in real time, Anthropic disabled both models for everyone, including every Amazon Bedrock customer. It's the first time Washington has reached straight into a live commercial deployment, which makes frontier-model access a policy lever, not just a vendor SLA. If your production stack hard-codes a single frontier model, add an abstraction layer and a tested fallback now — so the next directive or outage degrades your app instead of dropping it.
HOW TO READ THIS Read top to bottom: a new model enters a simulated deployment, 1.3M test conversations flow through it, misbehavior gets caught by the shield, and only the clean model reaches real deployment.
Replaying ~1.3M real, de-identified conversations through a candidate before ship turns safety testing from spot-checks into forecasting — and it already surfaced a novel 'calculator-hacking' failure.
HOW TO READ THIS Read top to bottom: AWS and QuEra join, set a Braket roadmap to 2028, then noisy physical qubits get error-corrected into a protected logical qubit on the Libra megaquop-class chip.
One of the most concrete timelines yet for error-corrected quantum you can actually rent — 'Libra' aims for a million operations across hundreds of logical qubits on Braket.
HOW TO READ THIS Read top to bottom: the two partners, the noisy qubits they started with, the correction loop that suppresses errors, and the resulting 800x drop confirmed in Nature.
Peer-reviewed Nature results show up to 800x lower logical error rates on trapped-ion qubits — the reliability floor quantum needs before it does useful work.
Open-source LLM engineering platform — evals, tracing, prompt management — the observability layer for anything you ship on top of a model.
Open-source RAG engine that pairs deep document understanding with agent workflows — production-grade retrieval without rolling your own.
Open multimodal agent stack that drives a real desktop UI — ByteDance's bet on agents that click, not just chat.
Official Chrome DevTools exposed as an MCP server, so coding agents can inspect, debug, and drive a live browser.
Indexes a codebase into a persistent knowledge graph over MCP — durable code intelligence instead of agents re-reading files every session.
Interconnects The case that we've entered the AGI era of AI governance — direct context for today's export-control move.
Latent Space GLM-5.2 claims the top frontend-coding slot — open-weight models keep closing on the frontier.
The Sequence Google DeepMind's first real crack at moving beyond next-token generation — early, but architecturally significant.