AI News — Friday, April 24, 2026

OpenAI has officially launched GPT-5.5, a significant step in its strategy to integrate AI capabilities into a comprehensive 'super app' experience.

OpenAI formally announced the release of GPT-5.5, highlighting its enhanced capabilities and setting a new benchmark for large language models.

OpenAI published the GPT-5.5 System Card, providing in-depth information on the model's safety measures, limitations, and advanced functionalities.

Google's decision to split its Tensor Processing Unit (TPU) into two chips indicates a strategic move to optimize hardware for increasingly complex agentic AI workloads.

Researchers introduce LLaDA2.0-Uni, a novel diffusion large language model designed to unify multimodal understanding and generation tasks, showing high community interest.

Google's AI Blog explains the architectural innovations and operational efficiencies of its TPUs, crucial for handling the escalating demands of modern AI applications.

Bret Taylor's new venture, Sierra, has acquired Fragment, an AI startup backed by Y Combinator, signaling significant investment and consolidation in the AI startup landscape.

A new research paper explores Near-Future Policy Optimization, a method aimed at improving the efficiency and effectiveness of AI policy learning in dynamic environments.

DR-Venus presents a breakthrough in developing frontier edge-scale deep research agents that can operate effectively with only 10K open data points, enhancing accessibility for resource-constrained environments.

A new paper delves into the critical issue of reward hacking in large models, analyzing its mechanisms, emergent misalignment, and the challenges it poses for AI safety.

ShadowPEFT introduces a new parameter-efficient fine-tuning method utilizing a shadow network, offering a more resource-friendly approach to adapt large language models.

OpenMobile proposes a framework for creating open mobile agents by synthesizing tasks and trajectories, pushing the boundaries of autonomous mobile AI.

DeVI explores physics-based dexterous human-object interaction through synthetic video imitation, advancing robotic manipulation and human-robot collaboration.

Google announced its first data center investment in Austria, signifying a strategic expansion of its global cloud infrastructure to support growing AI and data demands in the Alps region.

Noscroll is a new AI bot that promises to handle users' doomscrolling habits, curating and summarizing negative news to save time and mental energy.