AI News — Monday, May 25, 2026

10
OpenAI and Malta Partner to Bring ChatGPT Plus to All Citizens

OpenAI has announced a partnership with Malta to provide all Maltese citizens with free access to ChatGPT Plus, marking a significant step in national AI adoption and digital equity.

OpenAI Blogindustry
9
Introducing OpenAI for Singapore

OpenAI is expanding its global presence by launching operations in Singapore, aiming to foster AI innovation and collaboration within the Southeast Asian region.

OpenAI Blogindustry
9
Everyone's Talking About Gemini 3.5 Flash. The Real Story at Google I/O 2026 Was a Skill File.

While Gemini 3.5 Flash garnered attention at Google I/O 2026, the author argues that the introduction of a 'Skill File' for AI agents was the more significant and overlooked development.

Dev.toproduct
8
Unsupervised Process Reward Models

Researchers introduce a novel method for training reward models without human supervision, potentially streamlining the development of reinforcement learning from human feedback (RLHF) for AI systems.

Hugging Faceresearch
8
FlowLong: Inference-time Long Video Generation via Manifold-constrained Tweedie Matching

A new research paper presents FlowLong, a method enabling efficient inference-time generation of long videos by leveraging manifold-constrained Tweedie matching.

Hugging Faceresearch
7
SpaceDG: Benchmarking Spatial Intelligence under Visual Degradation

This paper introduces SpaceDG, a new benchmark designed to evaluate the robustness of AI models' spatial intelligence when faced with various forms of visual degradation.

Hugging Faceresearch
7
Everyone is navigating AI security in real time — even Google

The article highlights that even major tech companies like Google are actively and continuously developing strategies to address the evolving challenges of AI security.

TechCrunchindustry
7
Q-ARVD: Quantizing Autoregressive Video Diffusion Models

Q-ARVD proposes a method for quantizing autoregressive video diffusion models, significantly improving their efficiency and reducing computational requirements.

Hugging Faceresearch
7
Maestro: Reinforcement Learning to Orchestrate Hierarchical Model-Skill Ensembles

Maestro introduces a reinforcement learning framework for orchestrating hierarchical ensembles of models and skills, enhancing the capabilities of complex AI agents.

Hugging Faceresearch
7
Training Large Language Models to Predict Clinical Events

Researchers explore the application of large language models to predict clinical events, showcasing their potential to revolutionize healthcare diagnostics and patient management.

Hugging Faceresearch
6
KVServe: Service-Aware KV Cache Compression for Communication-Efficient Disaggregated LLM Serving

KVServe presents a novel approach to compress KV caches in large language models, leading to more communication-efficient and scalable disaggregated LLM serving architectures.

Hugging Faceresearch
6
I Ditched Cloud LLMs for Gemma 4 4B: A DevOps Engineer's 48-Hour Reality Check

A DevOps engineer shares their experience and insights after switching from cloud-based LLMs to the locally run Gemma 4 4B model, providing a practical perspective on on-device AI.

Dev.toopen-source
6
A new experiment brings better group meetings to Google Beam

Google Beam is introducing a new experimental feature designed to enhance group meeting experiences with improved AI-powered functionalities.

Google AI Blogproduct
6
Dissecting ThunderKittens, anatomy of a compact DSL for high-performance AI kernels

This article provides a deep dive into ThunderKittens, a compact domain-specific language (DSL) designed for developing high-performance AI kernels, offering insights into its architecture and benefits.

Lobste.rsopen-source
5
I spent 31 hours on the math behind TurboQuant so you don't have to

The author demystifies the complex mathematical underpinnings of TurboQuant, offering a simplified explanation to help developers understand and implement efficient quantization techniques for AI models.

Lobste.rsresearch