AI News — Thursday, June 25, 2026

10
OpenAI and Broadcom Unveil LLM-Optimized Inference Chip

OpenAI and Broadcom have partnered to introduce a new 'Jalapeno' inference chip specifically designed to optimize performance and efficiency for large language models.

OpenAI Blogproduct
9
Qwen-AgentWorld: Language World Models for General Agents

A new research paper introduces Qwen-AgentWorld, a framework utilizing language world models to enhance the capabilities of general AI agents, achieving strong performance across various benchmarks.

Hugging Faceresearch
8
NatureBench: Can Coding Agents Match the Published SOTA of Nature-Family Papers?

NatureBench evaluates the current state-of-the-art coding agents, questioning if they can autonomously replicate or surpass the performance reported in high-impact scientific publications.

Hugging Faceresearch
8
Listen Labs Raises $69M to Scale AI Customer Interviews After Viral Stunt

Listen Labs secured $69 million in funding following a viral billboard hiring campaign, aiming to expand its AI-powered platform for customer interviews and insights.

VentureBeatindustry
7
Europe Pushes Back on Washington's Chip War

European nations are reportedly challenging Washington's aggressive stance in the global chip war, signaling potential shifts in semiconductor supply chain dynamics and AI hardware access.

TechCrunchindustry
7
Cerebras Stock Plunges After Earnings, CEO Cites Misunderstood Margin Outlook

AI chip maker Cerebras Systems saw its stock drop significantly after its latest earnings report, with the CEO attributing the decline to a misunderstanding of the company's margin outlook.

TechCrunchindustry
7
RIFT-Bench: Dynamic Red-teaming For Agentic AI Systems

A new benchmark, RIFT-Bench, is introduced to dynamically red-team agentic AI systems, aiming to identify and mitigate potential vulnerabilities and failure modes in complex AI behaviors.

arXivresearch
6
Neuro-Symbolic Drive: Rule-Grounded Faithful Reasoning for Driving VLAs

Researchers propose Neuro-Symbolic Drive, a method that combines neural networks with symbolic rules to enable faithful and explainable reasoning for Vision-Language Agents in autonomous driving scenarios.

arXivresearch
6
OpenThoughts-Agent: Data Recipes for Agentic Models

OpenThoughts-Agent presents a collection of data recipes designed to improve the training and performance of agentic AI models, offering practical guidance for developers.

Hugging Faceresearch
6
AOHP: An Open-Source OS-Level Agent Harness for Personalized, Efficient and Secure Interaction

AOHP introduces an open-source, operating system-level agent harness designed to facilitate personalized, efficient, and secure interactions for AI agents.

Hugging Faceopen-source
6
Escaping the Self-Confirmation Trap: An Execute-Distill-Verify Paradigm for Agentic Experience Learning

This paper proposes an 'Execute-Distill-Verify' paradigm to help agentic AI systems avoid self-confirmation biases and improve learning from experience.

Hugging Faceresearch
5
Former Infosys Chief Launches New Startup to Challenge IT Services World

A former Infosys CEO has unveiled a new startup aiming to disrupt the traditional IT services industry, likely leveraging advanced AI and automation solutions.

TechCrunchindustry
5
Critique of Agent Model

This paper offers a critical analysis of current agent models, highlighting their limitations and suggesting directions for future research and development.

Hugging Faceresearch
5
Semantic Browsing: Controllable Diversity for Image Generation

Semantic Browsing introduces a novel approach to image generation that allows for controllable diversity by enabling users to semantically browse and refine generated outputs.

Hugging Faceresearch
5
FLAT: Feedforward Latent Triangle Splatting for Geometrically Accurate Scene Generation

FLAT proposes a new method using feedforward latent triangle splatting to achieve geometrically accurate and high-quality 3D scene generation.

Hugging Faceresearch