Sia Reads What Matters — Weekly AI Digest

NEWS2026-03-23

Anthropic refused to help the Pentagon spy on Americans — now a federal judge decides what happens next

The Anthropic v. Department of War saga reaches its inflection point. Senator Warren launched a formal investigation calling the supply-chain-risk designation 'retaliation' for Anthropic's refusal to allow Claude in mass surveillance and autonomous weapons. A federal hearing on March 24 will determine whether Anthropic gets preliminary relief while its lawsuit against the government plays out.

Anthropic's Pentagon Blacklist Heads to Court as Warren Calls It Retaliation

Anthropic's statement Warren investigation CNBC coverage

RELEASE2026-03-20

NVIDIA's Vera Rubin delivers 10x inference throughput per watt and Jensen sees $1T by 2027

GTC 2026's headline: seven new chips in full production powering the Vera Rubin platform — Vera CPU, Rubin GPU, NVLink 6, ConnectX-9, BlueField-4, Spectrum-6, and the integrated Groq 3 LPU. The NVL72 rack trains large MoE models with one-fourth the GPUs of Blackwell at one-tenth the cost per token. Jensen raised NVIDIA's revenue forecast to $1 trillion through 2027.

NVIDIA announcement GTC 2026 recap CNBC analysis

HOT TAKE2026-03-19

Perplexity is ditching MCP internally — tool schemas are eating the context window alive

At Ask 2026, Perplexity CTO Denis Yarats announced the company is moving away from MCP in favor of traditional APIs and CLIs. The core complaint: tool descriptions consume up to 72% of available context before the agent does any real work. Perplexity launched its own Agent API as an alternative — a single endpoint routing to models from OpenAI, Anthropic, Google, and others with built-in tools under one key.

Coverage Analysis

RELEASE2026-03-19

A mystery model topped OpenRouter's leaderboard for a week — turns out Xiaomi built it

An anonymous model called Hunter Alpha appeared on OpenRouter on March 11 with a trillion parameters and a million-token context window. Everyone assumed DeepSeek V4. It was actually MiMo-V2-Pro from Xiaomi's AI division, led by a former DeepSeek researcher. It ranks 8th globally on the Artificial Analysis Intelligence Index and is now partnering with five agent frameworks including OpenClaw and Cline for free developer access.

Xiaomi's Stealth Trillion-Parameter Model Unmasked as MiMo-V2-Pro

MiMo-V2-Pro Japan Times report

NEWS2026-03-21

Cursor doubled revenue in three months — and the market is already asking if IDEs are the wrong bet

Fortune's deep dive captures the tension: Cursor crossed $2B in annualized revenue and is in talks for a $50B valuation, yet the narrative is shifting from 'AI-enhanced IDEs' to 'autonomous coding agents.' Claude Code's $2.5B run rate and the viral 'Cursor is dead' discourse underscore the question — is the future the editor, or the agent that replaces it?

Fortune deep dive Bloomberg on $50B round

RESEARCH2026-03-18

DeepMind just defined what AGI actually means — and is paying $200K for you to help test it

Google DeepMind released a paper proposing a 10-ability cognitive taxonomy grounded in psychology and neuroscience for evaluating AI systems' progress toward AGI. The framework covers perception, attention, memory, learning, executive functions, reasoning, metacognition, language, action, and social cognition. They launched a $200K Kaggle hackathon to build evaluations for the five abilities with the largest gaps.

Google blog post The Register coverage

RESEARCH2026-03-17

China accounts for 41% of Hugging Face downloads — DeepSeek and Qwen are driving an open-model derivatives boom

Hugging Face's Spring 2026 State of Open Source report shows Chinese models have overtaken US models in Hub adoption for the first time. DeepSeek and Qwen drive the majority of downloads, with a fast-growing layer of quantized and fine-tuned derivatives. Western labs are responding with renewed urgency around GPT-OSS, OLMo, and Gemma.

Hugging Face report Analysis