AI News — Saturday, November 21, 2026

9
Google Search Upgraded with Gemini 4 Hydrogen for Enhanced Capabilities

Google announces a significant upgrade to its Search product, integrating the advanced Gemini 4 Hydrogen model to deliver more powerful and intelligent search results.

Google AI Blogproduct
9
AI Claims Solution to Riemann Hypothesis, Alongside Human Breakthrough

An AI system has reportedly claimed a solution to the 167-year-old Riemann hypothesis, coinciding with a human researcher also announcing a potential breakthrough, highlighting AI's growing capabilities in complex mathematics.

TechCrunchresearch
8
Google Introduces Gemini Decisions API for Fast, Calibrated Multimodal Decisions

Google launches the Gemini Decisions API, offering developers a new tool for making rapid and accurate multimodal decisions, expanding the practical applications of AI.

Google AI Blogproduct
8
Daniel Ek Backs AI Researcher's Fusion Power Startup

Spotify CEO Daniel Ek invests in an AI researcher's new startup focused on fusion power, signaling a significant venture capital interest in AI-driven energy solutions.

TechCrunchindustry
8
The Sleep Framework: A New Approach for LLMs to Retain Knowledge in Continual Learning

Google Research introduces 'The Sleep Framework,' a novel method designed to enable large language models to better retain knowledge and prevent catastrophic forgetting during continual learning processes.

Lobste.rsresearch
7
OpenAI Introduces MedChemBench to Aid Drug Discovery

OpenAI unveils MedChemBench, a new initiative aimed at leveraging AI to help scientists more efficiently identify promising drug candidates, accelerating pharmaceutical research.

OpenAI Blogresearch
7
Hebbia Improves Financial Research with OpenAI's GPT-6 Astra

Financial research firm Hebbia reports significant advancements in its operations by integrating OpenAI's latest large language model, GPT-6 Astra, demonstrating real-world enterprise application of cutting-edge AI.

OpenAI Blogindustry
7
Mind2Act World: A Diagnostic Benchmark for Embodied AI in Robotic Manipulation

Researchers introduce Mind2Act World, a new diagnostic benchmark designed to evaluate embodied AI's ability to integrate mind, physical action, and their composition in complex robotic manipulation tasks.

Hugging Faceresearch
7
Probing 'Conscious' Mind: Selective Access and Global Workspace in Vision-Language Models

New research explores the concept of 'consciousness' in Vision-Language Models by investigating selective access and global workspace mechanisms, shedding light on how these models process and integrate information.

Hugging Faceresearch
6
WeMM-Video: Advancing Open-Source Omni-Modal Video Understanding

WeMM-Video pushes the boundaries of open-source AI, offering a new framework for comprehensive omni-modal video understanding that integrates various data types for richer analysis.

Hugging Faceopen-source
6
Agentic Engineering of Smart Home Energy System with Antigravity and Gemini 4

A new project demonstrates the use of agentic engineering with Antigravity and Gemini 4 to create an intelligent smart home energy system, showcasing practical AI applications for energy management.

Dev.toproduct
6
Coda-DaVinci: Reimagining Video Generation with Code2Video

Coda-DaVinci introduces an innovative Code2Video approach, fundamentally reimagining how video content can be generated through programmatic instructions, offering new creative possibilities.

Hugging Faceresearch
5
Posterior Matching Distillation for Enhanced Model Training

A new research paper introduces Posterior Matching Distillation, a technique aimed at improving the efficiency and performance of model training by aligning posterior distributions.

Hugging Faceresearch
5
Depth Pro 2: Achieving Sharp and Temporally Consistent Metric Monocular Depth

Depth Pro 2 presents advancements in monocular depth estimation, delivering sharper and more temporally consistent metric depth maps from a single camera input, crucial for robotics and AR.

Hugging Faceresearch
5
The Mask Is Not the Object: Volumetric Supervision for 3D Gaussian Segmentation

This research proposes a novel method for 3D Gaussian segmentation using volumetric supervision, addressing challenges where traditional mask-based approaches fall short in complex 3D environments.

Hugging Faceresearch