AI News — Thursday, October 29, 2026
Microsoft announced it has generated $10 billion in revenue from its partnership with OpenAI, highlighting the significant commercial success of their collaboration.
New research introduces "The Trinity," a framework enabling multimodal AI agents to recursively self-improve their capabilities, pushing the boundaries of autonomous AI development.
Google is expanding its AI-powered search advertisements worldwide as the growth of its traditional ad business decelerates, signaling a major strategic shift towards AI monetization.
Microsoft disclosed Azure's revenue reached $33.2 billion, underscoring the massive scale and financial success of its cloud computing platform, heavily driven by AI services.
This paper presents a novel multiscale policy optimization technique designed to improve the performance of LLM agents in complex tasks requiring long-horizon planning and execution.
OpenAI and ServiceNow announced an expanded partnership aimed at integrating advanced AI capabilities into a broader range of enterprise workflows, accelerating AI adoption in businesses.
A new arXiv paper proposes a framework for human scientific endeavor in an era where AI increasingly surpasses human capabilities in scientific discovery, emphasizing human understanding and direction.
New research introduces Denoising Token Credit Assignment (DTCA), a method to improve reinforcement learning in diffusion language models by more effectively attributing credit to individual tokens.
This paper analyzes the computational and performance trade-offs associated with incorporating memory into AI agents, providing insights into its true value and cost.
OpenAI highlights Legora, a new venture focused on developing an agentic operating system specifically designed to automate and streamline various legal workflows.
VibeVideoBench is introduced as a new benchmark for rigorously evaluating the capabilities of multimodal AI agents in generating videos based on code-driven instructions.
This research presents PDRVLA, a novel vision-language-action model that incorporates physical dynamics representations to enable more realistic and effective interaction with the real world.
A practical guide offers insights and a benchmarking tool for developers to efficiently reduce the input token count for their AI agents, optimizing performance and cost.
The article discusses Mercury 3, an AI system demonstrating the capability to operate at speeds sufficient for real-time human-AI collaboration, enabling AI to remain actively involved in dynamic processes.
This paper investigates whether evolving an AI's memory during testing can yield better results for scientific discovery than traditional test-time training methods.