Live Feed
Microsoft Fairwater Goes Live Ahead of Schedule — $3.3B Wisconsin Campus with Hundreds of Thousands of NVIDIA GB200s
CEO Satya Nadella says Fairwater, Microsoft's flagship AI data center in Mount Pleasant, Wisconsin, is going live ahead of schedule. The $3.3B campus houses one of the densest concentrations of NVIDIA GB200 chips in the world, and Microsoft has already received approval to add 15 more buildings to the site.
OpenAI Codex Chronicle Captures Your Mac Screen to Build AI Context — Stores It Unencrypted, Skips EU
Chronicle, a research preview inside Codex Desktop for macOS, takes periodic screenshots, sends them to OpenAI servers for processing, and writes summaries to local unencrypted Markdown files. The system helps Codex remember active tasks without user prompting — and is explicitly excluded from the EU, UK, and Switzerland.
Alibaba Releases Qwen3.6 Max Preview Free — AA Score 52 Places It Atop Every Chinese Model
Alibaba released Qwen3.6 Max Preview today, scoring 52 on the Artificial Analysis Intelligence Index and ranking #2 among all tracked models. Free via Qwen Studio during preview, with paid API access coming via Alibaba Cloud. Beats Plus across six agentic coding and reasoning benchmarks.
Context.ai OAuth Breach Pivots Into Vercel — A Textbook AI Tool Supply Chain Attack
Attackers compromised Context.ai using Lumma infostealer malware, harvested OAuth tokens, replayed them to access a Vercel employee's Google Workspace, then pivoted into Vercel's internal systems. ShinyHunters has claimed credit. Mandiant and CrowdStrike are both on the case.
Berkeley RDI Backs AgentBeats With $1M+ Competition to Standardise Agent Evaluation Across 13 Domains
AgentBeats, a centralised agent benchmark registry built on the Agentified Agent Assessment (AAA) paradigm, launches with backing from Berkeley's Responsible Decentralized Intelligence lab and a $1M+ competition. The platform covers 13 agent domains including coding, healthcare, legal, DeFi, and agent safety, and integrates tau2-bench natively.
AI Medical Chatbots Hit 95% in Lab Tests — Then Drop to 35% When Real Patients Talk to Them
New research and a BBC investigation expose a systematic accuracy collapse in AI health chatbots: near-perfect performance on clean, structured medical cases gives way to 35% accuracy when real users describe symptoms the way real people actually do — incomplete, distracted, non-linear. A single phrasing change can flip advice from 'rest at home' to 'go to hospital now.'
Honor's Lightning Runs 50:26 — Humanoid Robot Breaks Human Half-Marathon World Record in Beijing
At the 2026 Beijing E-Town Humanoid Robot Half-Marathon, smartphone maker Honor's 'Lightning' robot completed 21.1 km in 50 minutes 26 seconds, breaking the human men's world record of 57:20. Last year's winning robot took over 2 hours 40 minutes. Twelve months of progress, measured in asphalt.
DeepSeek V4 Is Weeks Away: 1 Trillion Parameters, Huawei Chips, Apache 2.0
Multiple credible sources including Reuters and The Information point to a late-April launch for DeepSeek V4 — built entirely on Huawei Ascend hardware, estimated at 1 trillion parameters with 37 billion active per token, and expected under Apache 2.0. The model has already missed two earlier windows.
Gemma 4's 2B Edge Model Runs Fully Offline on iPhone — 1.5 GB, Apple Neural Engine, No Cloud Required
Google's Gemma 4 E2B and E4B edge variants are running on-device on iPhones via the Apple Neural Engine, distributed through Locally AI and Google's own AI Edge Gallery. The 1.5 GB quantized footprint fits comfortably in a modern iPhone's memory budget, with all inference staying on-device.
xAI Drops Grok 4.3 Beta at 500B Parameters With Zero Fanfare — 1T Version Due April 22
Grok 4.3 appeared in the model selector on April 17 with no blog post, no press release, and no announcement. Elon Musk confirmed the beta runs 500 billion parameters. The full 1 trillion parameter model completes training around April 22-23.
LLMs De-Anonymize Online Users at 68% Match Rate — 9 in 10 Guesses Correct
A new paper demonstrates that LLMs can link anonymous accounts to real identities across platforms with 68% recall and 90% precision, shattering the assumption that pseudonyms protect privacy when public writing is available.
NSA Deploys Anthropic Mythos for Cyber Work While Its Own Parent Department Calls Anthropic a Supply Chain Risk
The National Security Agency is using Claude Mythos Preview for offensive and defensive cyber operations despite the Department of Defense — which oversees the NSA — formally flagging Anthropic as a supply chain risk. The split creates a public contradiction at the centre of the US government AI security posture.
NVIDIA Ships Ising: Open AI Models for Quantum Error Correction, 2.5x Faster Than pyMatching
NVIDIA launched Ising, the first open-source AI model family purpose-built for quantum computing infrastructure. Two models — a vision-language calibrator and a 3D CNN error decoder — tackle the two engineering bottlenecks blocking practical quantum hardware. IONQ surged 50%, QUBT 30% on the announcement.
Cross-Datacenter LLM Inference Is Now Viable: Prefill-as-a-Service Delivers 54% Throughput Gain
A new paper from researchers running a 1-trillion-parameter hybrid model shows that splitting LLM inference across datacenters — routing long prompts to a remote prefill cluster while keeping decode local — delivers 54% higher throughput over local-only serving and 32% over naive heterogeneous setups. The enabling factor is smaller KV cache in hybrid-attention architectures.
Tinder and Zoom Adopt Iris Scanning as AI Makes Faces Untrustworthy
World — formerly Worldcoin — has signed Tinder and Zoom as its first major platform partners for World ID, iris-based proof of humanity. The move signals that biometric personhood verification is becoming a practical layer on apps flooded with AI-generated fakes.