OpenAI Discloses AI Policy Stance on RegulationAlphabet Plans $80B AI Infrastructure ExpansionOpenAI Breaks Ground on 1GW Data Center in MichiganxAI Unveils Grok Imagine Video Agent BreakthroughMAVEN Improves Agent Tool GeneralizationUnicorn Model Breaks Dimension Limits in Time Series ForecastingEHRBench: LLM Clinical Decision BenchmarkNumLeak Exposes Data Leakage in Foundation ModelsLinear Ensembles Break LLM Watermark DetectionSLAT Trims Redundancy in LLM ReasoningAlibaba's Qwen 3.7 Plus on Vercel AI GatewayOpenAI Frontier Models Now Available on AWSVercel Build Machines Add Memory ProtectionOpenAI Codex 0.136.0: TUI Adds Link ArchivingOpenClaw 2026.6.1-beta.2: Stable Tool CallsStanford Releases AI Agent Guidelines for CS336DuckDuckGo's No-AI Search Traffic SurgesFlorida Sues OpenAI and Sam AltmanHackers Compromised Instagram via Meta AI BotHackers Hijack Instagram Accounts via Meta's AI BotOpenAI Discloses AI Policy Stance on Regulation

Alphabet plans an $80 billion equity raise to expand AI infrastructure and computing. The record funding will support next-gen model training and strengthen its cloud/AI compute market position.

OpenAI has started construction on a 1GW data center in Michigan as part of the Stargate initiative. The project aims to expand AI access, create jobs, and support communities, building critical infrastructure for the Intelligence Age.

xAI’s tech lead Ethan He reveals how Grok Imagine was built in 3 months. The model unifies Videogen and world models for video generation and understanding. He argues it’s underrated and shows why video agents are the future.

arXiv paper MAVEN addresses LLMs’ poor generalization in agentic tool-calling. The method improves cross-environment performance by 30% and enhances multi-tool collaboration reliability through optimized reasoning strategies.

Unicorn solves a core trade-off in time series forecasting: channel-independent models scale but ignore dependencies, while channel-dependent models are dimension-bound. By using universal correlation modeling, it breaks through dimensional limitations and shows significant performance gains on high-dimensional data. Developers can directly apply this architecture to boost prediction systems, especially for finance, meteorology and other multivariate scenarios.

EHRBench introduces an automated benchmark for evaluating LLMs on clinical decision-making tasks using electronic health records. It measures diagnosis, treatment selection, and health outcome prediction capabilities.

Researchers found that public numeric benchmarks appear during pretraining, risking evaluations that measure memorized recall rather than true skill. NumLeak provides a framework to detect this data leakage via API probes.

Research reveals that watermarking for AI-generated text fails when users access multiple models. The statistical signatures embedded in watermarks are trivially washed away by distributional perturbations in multi-model environments.

SLAT introduces segment-level adaptive trimming to eliminate structural redundancy in LLM reasoning chains. The method maintains accuracy while significantly improving computational efficiency for large reasoning models.

Alibaba’s Qwen 3.7 Plus is now on Vercel AI Gateway, offering a unified multimodal agent foundation. Supports GUI/CLI operations, coding workflows with full-modality inputs, and visual agent tasks like perception and reasoning.

OpenAI’s frontier models and Codex are now generally available on AWS, allowing enterprises to build with OpenAI through existing AWS environments and workflows. This accelerates deployment from evaluation to production.

Vercel’s build machines now monitor memory usage and auto-scale resources. Memory-intensive builds won’t be downgraded, and near-OOM builds get upgrades. This prevents out-of-memory failures for developers.

OpenAI Codex 0.136.0 updates TUI with clickable web links and auto-conversion of cramped tables to key-value records. Adds CLI session archiving and better recovery from interrupted tool calls, improving developer workflow.

OpenClaw 2026.6.1-beta.2 improves recovery from interrupted tool calls, session bindings, and media delivery. Enhances stability across Telegram, WhatsApp, and iMessage channels with better CLI runtime support.

Stanford’s CS336 course has published official guidelines for using AI agents like Claude in assignments. The document specifies allowed use cases and limitations, receiving 311 points on GitHub.

DuckDuckGo has made its no-AI search engine more accessible, leading to a surge in traffic. The platform, which promises unaltered search results without AI rewriting, is gaining users frustrated by AI-filtered results.

Florida has sued OpenAI and CEO Sam Altman for failing to adequately disclose AI risks, violating state consumer protection laws. The lawsuit seeks to stop deceptive practices and seeks damages, reigniting AI regulatory debates.

Hackers tricked Meta’s AI support bot into granting access to high-profile Instagram accounts. The attack involves simple conversation prompts linking target accounts. Multiple sources confirm the vulnerability; Meta hasn’t commented.

Attackers exploited Meta’s AI support bot to seize Instagram accounts by bypassing security verification. The vulnerability allowed unauthorized account takeovers, with Meta yet to provide a timeline for fixes.