AI platforms are absorbing adjacent layers while regulators and product tests expose the resulting control points. OpenCV is widening into an edge runtime, the UK is separating conventional indexing from AI summaries, Nvidia is externalizing moderation policy, and Gemini Spark shows how connected data can surprise the person it describes.
1. OpenCV 5 rewrites its DNN engine and adds LLM support
OpenCV 5.0 ships a rewritten deep-neural-network engine, more than 80% ONNX coverage, built-in LLM and vision-language-model support, a new hardware abstraction layer, and expanded 3D vision tooling. Optimized paths cover Intel, Arm, Qualcomm, and RISC-V hardware, with native GPU support planned.
The rewrite moves OpenCV toward a portable edge-AI runtime, but native GPU execution remains planned. Current accelerator-heavy deployments therefore face a split between broad ONNX coverage on supported paths and a second runtime for missing execution targets.
Sources: Phoronix on the OpenCV 5.0 release · OpenCV 5.0 release repository
2. UK regulator orders an opt-out path for Google's AI summaries
Britain's Competition and Markets Authority ordered Google to let publishers exclude material from AI-generated search summaries without disappearing from conventional search. The measure addresses a central bundling problem: crawling for indexing and using the same work to generate substitutive answers had been offered as one choice.
The order separates permission to index a page from permission to use it in a substitutive answer. That distinction forces Google to preserve conventional ranking while honoring a second serving-time control over AI summaries.
Sources: Associated Press on the UK order
3. Nemotron 3.5 makes multimodal safety policy configurable
Nvidia released Nemotron 3.5 Content Safety, a 4-billion-parameter model based on Gemma 3. It evaluates prompts, optional images, and responses together; supports custom policy text and 12 trained languages; and can return either a low-latency verdict or a verdict with a reasoning trace.
Nvidia reports roughly 85% average accuracy on its selected benchmarks and publishes both weights and a safety dataset. The vendor controls the evaluation, while the model's custom-policy interface makes policy interpretation itself a measurable source of error across languages.
Sources: Nvidia's Nemotron 3.5 technical release
4. Gemini Spark reveals the privacy price of connected context
A Verge test of Google's Gemini Spark produced a detailed family itinerary using information the reviewer had not supplied in the prompt, including relatives' names, children's ages, a dog's name, food preferences, and concert tickets. Spark drew on connected Google data and failed only when Airbnb blocked a booking attempt.
The itinerary joined email, identity, purchase history, and family details before the reviewer supplied them explicitly. That surprise is the product tension: connected context reduces prompting effort while making the retrieval boundary invisible at the moment personal data enters an answer.
Sources: The Verge's Gemini Spark test · Google's May AI product roundup