AI News — Sunday, November 15, 2026
A new research paper explores whether AI agents learn the task itself or merely the verifier's criteria, proposing a re-evaluation of self-improvement methods in embodied AI.
Google announces Gemini Omni 1.2 Flash, a new model designed for faster and more cost-effective processing of longer video clips, enhancing its multimodal capabilities.
WorldDraft introduces a novel concept where video models are explored for their potential to render entire virtual worlds based on code, similar to how code builds software.
TechCrunch reports on questions surrounding Nvidia's security team and a petition signed by prominent tech executives, including Elon Musk, calling for Jensen Huang's reinstatement, indicating potential internal turmoil or external pressure.
New research delves into 'unbundling' multimodal vision encoding, aiming to improve how AI models process and understand information from diverse visual inputs.
A developer shares practical advice on setting up AI coding agents with '10 Lines' of instructions to ensure safe and effective interaction with existing code repositories.
A new paper introduces 'Factored One-sided Adam-Moment (SOAP)', an optimization algorithm promising practical and scalable performance for AI model training.
TechCrunch compiles a running list of AI projects and startups that have ceased operations, highlighting the high-risk nature and competitive landscape of the AI industry.
OpenAI shares new customer success stories, demonstrating how companies like Shield and Amplitude are leveraging OpenAI's models to streamline workflows, improve product development, and double revenue.
Researchers introduce 'Interactive Visual Gym,' a new framework for evaluating visual interaction models by simulating users within interactive virtual environments.
Developers discuss the limitations of AI coding agents, noting how they can omit crucial details in summaries and contribute to a 'review bottleneck' as code generation becomes cheap but judgment remains critical.
UniSpark proposes a hybrid attention mechanism to advance multimodal understanding and generation in AI models, improving their ability to process and create content across different data types.
New research explores 'Evolving Infra,' a method for agent-adaptive LLM serving at scale, addressing the infrastructure challenges of deploying large language models for dynamic AI agents.
A developer shares how they integrated AI to make Jira proactively check for completion, preventing bugs and improving the reliability of 'done' statuses in software development.
This paper investigates 'Who Remembers What?' by proposing a multi-granularity memory extraction technique to improve understanding and interaction in complex multi-user, multi-agent AI systems.