AI News — Saturday, July 18, 2026

9
VideoChat3: Fully Open Video MLLM for Efficient and Generalist Video Understanding

VideoChat3 introduces a new fully open-source video Multi-modal Large Language Model (MLLM) designed for efficient and generalist video understanding, offering significant advancements in the field.

Hugging Faceopen-source
9
Databricks hits $188B valuation, extending its run as AI’s favorite second act

Databricks has achieved a staggering $188 billion valuation, solidifying its position as a key player in the AI ecosystem by providing essential data and analytics infrastructure.

TechCrunchindustry
9
The agent security gap: 54% of enterprises have already had an AI agent incident, and most still let agents share credentials

A new report reveals a significant 'agent security gap,' with over half of enterprises experiencing an AI agent-related security incident and many still allowing agents to share sensitive credentials.

VentureBeatindustry
8
LongStraw: Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget

Researchers have developed 'LongStraw,' a new method enabling Reinforcement Learning (RL) to handle contexts exceeding 2 million tokens while operating within a fixed GPU budget, addressing a major limitation in long-sequence processing.

Hugging Faceresearch
8
SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning

The SEED framework proposes a novel approach for agentic reinforcement learning that uses self-evolving on-policy distillation to improve learning efficiency and agent performance.

Hugging Faceresearch
8
The AI compute gap: Enterprises are buying infrastructure faster than they can measure what it costs

Enterprises are rapidly acquiring AI compute infrastructure, creating a 'compute gap' where they struggle to accurately measure and manage the associated costs.

VentureBeatindustry
7
SearchOS-V1: Towards Robust Open-Domain Information-Seeking Agent Collaboration

SearchOS-V1 introduces a new framework aimed at fostering robust collaboration among open-domain information-seeking agents, enhancing their ability to work together effectively.

Hugging Faceresearch
7
Vertu wants executives to pay $6,880 for an AI agent — here’s how it actually performs

Vertu is marketing a high-priced AI agent for executives at $6,880, prompting scrutiny into its actual performance and value proposition for business leaders.

TechCrunchproduct
7
Experiments with On-device AI — What building on Gemini Nano actually teaches you

This article shares practical insights and lessons learned from developing applications using Google's Gemini Nano, highlighting the realities of on-device AI implementation.

Dev.toproduct
7
BadWAM: When World-Action Models Dream Right but Act Wrong

A new paper explores the phenomenon of 'BadWAM,' where World-Action Models accurately predict outcomes but fail to execute correct actions, highlighting a critical challenge in agentic AI.

Hugging Faceresearch
7
OriginBlame: Record- and Token-Level Data Provenance for AI Training Datasets

OriginBlame proposes a novel system for tracking data provenance at both record and token levels within AI training datasets, crucial for transparency and accountability.

arXivresearch
7
SPINE: Bridging the Cyber-Physical Gap with Agentic AI

SPINE presents an innovative approach to bridge the gap between cyber and physical systems using agentic AI, promising more integrated and intelligent control over real-world environments.

arXivresearch
7
The AI context gap: Enterprise AI organizations have a trust problem, not a retrieval problem — and most are still building the fix

Enterprise AI organizations face a 'context gap' rooted in trust issues rather than retrieval failures, with many still working to implement solutions to build confidence in AI outputs.

VentureBeatindustry
6
RoboTTT: Context Scaling for Robot Policies

RoboTTT introduces a method for context scaling in robot policies, enabling more complex and adaptable behaviors for robotic systems by handling larger contextual information.

Hugging Faceresearch
6
Create, edit and star in videos with two Google Vids updates

Google Vids has received two new updates, empowering users to create, edit, and feature themselves in videos more easily, leveraging AI for enhanced content creation.

Google AI Blogproduct