AI News — Thursday, April 2, 2026
Anthropic faced a major incident, accidentally issuing DMCA takedowns for thousands of GitHub repositories while attempting to remove its recently leaked Claude source code.
Speculation mounts regarding the recent Claude code leak, questioning whether it was an accident, a display of incompetence, or a calculated public relations maneuver by Anthropic.
Cognichip secured $60 million in funding to pursue its ambitious goal of using artificial intelligence to design the very chips that power AI systems.
New research introduces FIPO, a policy optimization method designed to elicit deeper reasoning capabilities in AI models through future-KL influence.
Meta's substantial investment in natural gas to fuel its AI infrastructure raises environmental concerns, with its energy demands potentially equivalent to powering an entire state like South Dakota.
CARLA-Air presents a novel unified infrastructure that allows for the simulation of drones within the CARLA world, advancing research in air-ground embodied intelligence.
LongCat-Next proposes a method for lexicalizing different modalities as discrete tokens, potentially leading to more efficient and powerful multimodal AI models.
Gradient Labs is leveraging OpenAI technology to offer every bank customer an AI-powered account manager, aiming to revolutionize financial services.
Project Imaging-X provides a comprehensive survey of over a thousand open-access medical imaging datasets, crucial for developing foundation models in healthcare AI.
Lingshu-Cell introduces a generative cellular world model capable of transcriptome modeling, paving the way for the creation of virtual cells in biological research.
Google AI is developing a new satellite imagery map to aid in the protection of Brazil's forests by monitoring and detecting deforestation.
OpenAI shares its vision and strategic initiatives for rapidly advancing the next generation of artificial intelligence technologies.
GEMS proposes a new framework for agent-native multimodal generation, integrating memory and skills to enhance AI agent capabilities.
VGGRPO presents a novel approach for video generation that utilizes a 4D latent reward mechanism to achieve more world-consistent and realistic video outputs.
OpenAI has released updated policies and safeguards for GPT models to assist developers in creating safer AI experiences specifically for teenage users.