Brief — March 31, 2026

Library Blog Radar Briefs Intelligence About

Apple Intelligence Accidentally Launches in China — Then Goes Dark

Apple Intelligence rolled out to iPhones in China in what Bloomberg's Mark Gurman confirmed was an "error." Beijing law requires Apple to partner with local companies (Alibaba) to power AI features — so Apple quickly pulled it offline. The real story: **the pressure to launch everywhere, fast, is creating governance failures at the OS level.** China's AI market is the contested territory nobody talks about openly. Apple's misstep reveals just how

Microsoft Copilot Cowork Goes Live — With Claude Doing the Editing

Microsoft launched Copilot Cowork via its Frontier Program. The standout detail: **GPT drafts the research, then Claude reviews it for accuracy.** This is the first major commercial deployment of a multi-model pipeline where competing AI labs are embedded in the same enterprise product. Also ships: an improved Researcher agent and a Critique feature. The era of single-model enterprise AI is over. The future is model orchestration — and Microsoft

Google Discloses Quantum Vulnerabilities in Crypto — Responsibly

Google published a responsible disclosure on quantum vulnerabilities affecting widely-used cryptographic schemes in blockchain and cryptocurrency infrastructure. The post reveals how current quantum hardware could begin threatening specific elliptic curve implementations sooner than anticipated. **This is the Q Day drumbeat getting louder.** Google moves carefully — publishing this means the threat is real enough to warrant warning the industry,

Ollama Now Powered by MLX on Apple Silicon — Local AI Accelerates

Ollama's new preview switches its Apple Silicon backend to **MLX**, Apple's machine learning framework. Practically: running local LLMs on M-series Macs just got meaningfully faster and more efficient. Combined with Apple's deal to use Gemini for training smaller on-device models, a picture emerges: **local AI inference is becoming viable for serious use cases.** The cloud-first assumption in AI is quietly eroding. > Source: Ollama Blog / HN (Mar

Axios NPM Package Compromised — Remote Access Trojan Shipped

Axios, the most popular HTTP client in JavaScript (used by millions of projects), was **compromised on NPM with malicious versions that drop a remote access trojan.** With 343 points and 94 comments on HN this morning, this is already a major supply chain security incident. Every Node.js project pulling Axios should audit their lock files immediately. The software supply chain is the new attack surface. > Source: StepSecurity / HN (Mar 31)

Nuclear Watchdog Gutted as Big Tech Pushes Reactor Revival

Silicon Valley is championing next-gen nuclear reactors to power data centers — but the Nuclear Regulatory Commission has lost 400+ staff, largely those working on safety. Former NRC chair Allison Macfarlane: *"The regulator is no longer an independent regulator — we do not know whose interests it is serving."* The AI energy demand story has a dark undercurrent: **the safety infrastructure meant to oversee the power infrastructure is being disman

Blog Pitches

The Infrastructure Wars Nobody Is Watching — Apple vs China distribution, nuclear safety gutted for data centers, quantum threatening the cryptographic layer. The real AI competition isn't about model benchmarks — it's about who controls the stack beneath the model. [Pillar: Tech Sovereignty]

Multi-Model Pipelines Are the New Normal — What Copilot Cowork Reveals — When Microsoft ships GPT + Claude in the same product, the single-model era is over. What does enterprise AI look like when it's orchestration, not a single vendor? Who wins? [Pillar: AI & Education / Creative Intelligence]

Local AI Is Quietly Getting Serious — Ollama on MLX, Apple distilling Gemini for on-device use. The cloud-first assumption in AI is eroding. What happens when inference moves to the edge? [Pillar: Frontiers / The New Playbook]

When Your NPM Package Delivers Malware — The Axios compromise is a case study in supply chain vulnerability. Every developer is one malicious package away from owning their users. What does secure software dependency look like in 2026? [Pillar: Tech Sovereignty]

LessWrong

The safety conversation is shifting from — will AI go rogue" to "how does misalignment emerge structurally from how we train it." Reward hacking as a source of emergent misalignment is the technical argument. "Good citizen vs good assistant" is the values argument. Both circling the same problem: the default optimization target for AI is not aligned with human flourishing.