AI News — Wednesday, March 25, 2026

10
OpenAI’s Sora Shuts Down After Being Deemed 'Creepiest App'

OpenAI's experimental Sora app, previously known for its advanced capabilities, has been abruptly shut down, with TechCrunch reporting it was considered the 'creepiest app' on users' phones.

TechCrunchproduct
9
Kleiner Perkins Secures $3.5B to Double Down on AI Investments

Venture capital firm Kleiner Perkins has raised $3.5 billion in fresh capital, signaling a significant strategic pivot to invest heavily in the burgeoning AI sector.

TechCrunchindustry
8
OpenAI Introduces New Policies to Help Developers Build Safer AI for Teens

OpenAI has announced new guidelines and safeguards aimed at assisting developers in creating safer AI experiences, particularly for teenage users, emphasizing responsible AI development.

OpenAI Blogindustry
8
OpenAI Provides Update on the OpenAI Foundation's Structure and Mission

OpenAI has released an update regarding the operational status and strategic direction of the OpenAI Foundation, clarifying its role within the broader organization.

OpenAI Blogindustry
8
Google Research Unveils TurboQuant for Extreme AI Model Compression

Google Research introduces TurboQuant, a groundbreaking technique promising to redefine AI efficiency through extreme compression of models, enabling faster and more resource-efficient AI applications.

Lobste.rsresearch
7
ChatGPT Integrates New Features to Power Product Discovery

OpenAI is enhancing ChatGPT with new functionalities designed to improve product discovery, allowing users to find and explore products more effectively through the AI assistant.

OpenAI Blogproduct
7
Omni-WorldBench: A New Standard for Comprehensive World Model Evaluation

Researchers introduce Omni-WorldBench, a novel benchmark designed for comprehensive, interaction-centric evaluation of world models, aiming to standardize performance assessment in complex AI environments.

Hugging Faceresearch
7
Speed by Simplicity: A Single-Stream Architecture for Fast Audio-Video Generative Models

A new research paper proposes 'Speed by Simplicity,' a single-stream architecture that significantly accelerates audio-video generative foundation models, improving efficiency without sacrificing quality.

Hugging Faceresearch
6
AI Agents Are Becoming API's Biggest Consumers: Do They Prioritize Good Design?

An article explores the growing trend of AI agents as primary consumers of APIs and raises critical questions about whether these agents inherently value or contribute to good API design.

Dev.toindustry
6
LongCat-Flash-Prover: Advancing Formal Reasoning with Agentic Tool-Integrated Reinforcement Learning

A new paper introduces LongCat-Flash-Prover, an innovative system that enhances native formal reasoning capabilities through the integration of agentic tools and reinforcement learning.

Hugging Faceresearch
6
OpenResearcher: An Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis

OpenResearcher presents a fully open-source pipeline designed to synthesize long-horizon deep research trajectories, aiming to automate and accelerate scientific discovery.

Hugging Faceopen-source
5
Dev.to User Builds Spotify Playlist Generator with Claude AI

A developer on Dev.to shares their experience building a functional Spotify playlist generator, showcasing the practical application of Anthropic's Claude AI for creative tasks.

Dev.toproduct
5
AgenticGEO: A Self-Evolving Agentic System for Generative Engine Optimization

New research introduces AgenticGEO, a self-evolving agentic system designed to optimize generative engines, promising advancements in the efficiency and performance of AI content creation.

arXivresearch
5
ProMAS: Proactive Error Forecasting for Multi-Agent Systems Using Markov Transition Dynamics

ProMAS is a new research development that enables proactive error forecasting in multi-agent systems by leveraging Markov Transition Dynamics, enhancing reliability and robustness.

arXivresearch
5
Domain-Specialized Tree of Thought through Plug-and-Play Predictors

Researchers propose a method for creating domain-specialized Tree of Thought models by integrating plug-and-play predictors, allowing for more targeted and efficient AI reasoning in specific fields.

arXivresearch