AI News — Friday, September 11, 2026
This paper introduces a new approach to scale automatic research agents by integrating world models, demonstrating significant advancements in autonomous scientific discovery.
OpenAI has temporarily halted new Pro subscriptions, citing overwhelming demand for its latest AI model, Astra, indicating rapid user adoption and growth.
Nvidia CEO Jensen Huang projects an impressive 70% growth for the company next year, driven by continued demand for AI infrastructure and specialized hardware.
Researchers propose AgentGrad, a novel method for optimizing prompts in multi-agent systems using intervention-guided techniques, enhancing collaboration and performance.
A new technical report explores the development of latent space language models using next concept prediction, aiming for more efficient and powerful AI architectures.
A new paper introduces a programmable world model, offering a flexible framework for AI agents to understand and interact with complex environments.
OpenAI announces the launch of ChatGPT for Financial Services, a specialized version designed to meet the unique demands and regulatory requirements of the finance industry.
A new article argues that AI has surpassed most human software developers in coding proficiency, highlighting the rapid advancement and impact of AI in programming.
SenseNova-U1.5 is introduced as a step towards native unified visual intelligence, aiming to integrate various visual tasks into a single, cohesive AI system.
This paper presents T1, a new reinforcement learning framework for terminal agents designed to tackle complex, long-horizon tasks more effectively.
SWE-Bench Pro Verified is introduced as a robust and reliable benchmark for evaluating the performance of AI agents in software engineering tasks.
SAEScientist-Bench explores the capability of AI agents to autonomously conduct research in SAE (Sparse Autoencoder) interpretability, pushing the boundaries of AI-driven scientific discovery.
A new protocol, Discovery Certification, is proposed to rigorously audit AI research agents, emphasizing that raw scores alone are insufficient to validate scientific discovery.
This research investigates the effectiveness of subagents versus agent skills in executing reusable knowledge for long-horizon agentic tasks, aiming to improve AI task completion.
A new paper explores adaptive entangled game modules as a foundational component for Artificial General Intelligence, suggesting a novel path for developing more versatile AI systems.