Google DeepMind acquihires Contextual AI in $80M licensing deal

DeepMind recruited more than 20 researchers from Contextual AI, including co-founder and CEO Douwe Kiela, under a non-exclusive licensing arrangement valued between $80 million and $90 million. The deal follows Google's earlier acquihires of Windsurf and Character.AI talent and continues a pattern designed to expand DeepMind's research capacity without triggering formal antitrust review.

Meta opens Ray-Ban Display glasses to third-party developers

Meta announced that third-party developers can now build display-enabled apps and Neural Band controller experiences for its $800 Ray-Ban Display glasses, expanding well beyond the handful of Meta-made apps at launch. The platform supports iOS, Android, and web-backed apps, with an emphasis on information overlays and real-time data displays suited to the monocular field of view.

OpenAI opens self-serve Ads Manager to US advertisers in beta

OpenAI launched a beta self-serve Ads Manager letting US businesses directly buy, manage, and measure ChatGPT ad campaigns, adding cost-per-click bidding alongside the existing CPM model. The platform integrates with agency holding companies including Dentsu, Omnicom, Publicis, and WPP, and the company is reportedly targeting $2.5 billion in ad revenue this year.

Anthropic ships Claude Opus 4.7 Fast Mode and lands PwC deal

Anthropic added Fast Mode in research preview for Claude Opus 4.7, delivering significantly faster output token generation at premium pricing—mirroring the existing Opus 4.6 fast mode but applied to its newest flagship. Separately, PwC signed on to deploy Claude for building technology, executing deals, and reinventing enterprise functions for clients, extending Anthropic's run of large professional-services wins this month.

Claude Opus 4.7 Jumps from 54% to 98.5% on Visual Computer-Use Benchmark

XBOW's autonomous penetration-testing team reported Claude Opus 4.7 scores 98.5% on their visual-acuity benchmark, up from 54.5% for Opus 4.6 — a near-doubling on the metric that gates reliable computer-use agents. Anthropic's model card confirms the gain and notes Opus 4.7 is the first release where cyber capabilities were deliberately reduced relative to Claude Mythos Preview, meaning the visual-acuity jump is happening in a safety-constrained model, not the lab's most capable one.

Frontier AI Finds Majority of Palo Alto Networks' Bugs Across 130 Products

Palo Alto Networks' May 'Patch Wednesday' release marks the first time the majority of findings in their security advisories came from frontier AI models scanning code — covering all 130+ products across three platforms. Their testing found three weeks of model-assisted analysis matched a full year of manual penetration testing in coverage, and that attack-cycle time from initial access to exfiltration has collapsed to as little as 25 minutes in AI-assisted scenarios.

HKUST's TideGS trains billion-primitive 3D scenes on a single GPU

TideGS, from HKUST's Sponge Computing Lab, enables training 3D Gaussian Splatting models with over one billion primitives on a single GPU by tiering parameters across an SSD-CPU-GPU hierarchy through block-virtualization, asynchronous pipelining, and differential streaming. Previously, scenes at this scale required multi-GPU clusters; TideGS eliminates that constraint without sacrificing gradient fidelity. The technique has direct implications for large-scale scene reconstruction in robotics and simulation pipelines.

NVIDIA Rubin platform claims 10x inference token cost cut over Blackwell

NVIDIA's Vera Rubin platform, entering H2 2026 deployment, delivers up to 10x reduction in inference token cost and 4x fewer GPUs to train MoE models versus Blackwell, with 3.6 ExaFLOPS FP4 compute and 22 TB/s HBM4 memory bandwidth per GPU. AWS, Google Cloud, Microsoft Azure, and CoreWeave are among the first cloud providers deploying Rubin-based instances. A new Inference Context Memory Storage Platform with BlueField-4 is also included to accelerate agentic AI reasoning workloads.

Trump delays AI model-vetting executive order after last-minute objections

President Trump postponed signing an executive order that would have created a voluntary framework giving the federal government up to 90 days of pre-release access to frontier AI models for cybersecurity review. The signing had been scheduled for Thursday and was abandoned hours before the ceremony after Trump told reporters he 'didn't like certain aspects' of the order. The pause leaves unresolved how the administration will respond to frontier models like Anthropic's Mythos, which officials have flagged for its ability to identify and exploit security vulnerabilities, marking a striking moment of internal friction for an administration that has otherwise championed a hands-off approach to AI.

Connecticut governor set to sign omnibus AI law covering employment, chatbots, and frontier developers

Connecticut's SB 5—the Connecticut Artificial Intelligence Responsibility and Transparency Act—cleared the legislature 131–17 in the House and 32–4 in the Senate on May 1, and Governor Ned Lamont has confirmed he will sign it. The 71-page law requires disclosure to workers when AI is a 'substantial factor' in employment decisions, bans automated decision-making as a defense against discrimination claims, imposes conduct rules on AI companion chatbots (including strict child-specific prohibitions), and mandates whistleblower protections for employees at frontier developers training models above 10²⁶ compute operations. Connecticut joins California, New York, and Colorado as a state legislating substantively on AI even as a federal preemption effort moves through the House.