AI News — Monday, August 10, 2026
Anthropic is making Claude Code's auto mode a default feature and has launched Cowork, a new desktop agent for Claude that operates directly on user files without requiring coding.
Nous Research has released NousCoder-14B, an open-source coding model that directly competes with commercial offerings like Anthropic's Claude Code, offering a free alternative for developers.
Hedge fund Situational Awareness, despite its own struggles, has made a significant $400 million investment in Source Foundry, a startup focused on developing AI chips.
A new research paper introduces PaDoc, a method for document parsing that uses layout-grounded parallel decoding to improve efficiency and accuracy.
CalibForge presents a novel approach to adversarial solver calibration, aiming to enhance the scalability and performance of learnable terminal tasks in AI systems.
Researchers propose KVAE, a new family of tokenizers designed to improve the performance and flexibility of multimodal generative models.
EffectLearner is a new system that leverages world-aware object-effect reasoning to enable realistic and efficient object removal in real-world videos.
Historian Jill Lepore argues that Silicon Valley's flawed understanding of science fiction is leading to technological developments that undermine democratic principles.
Vaya is introduced as an AI-powered loan advisor that uniquely focuses on assessing a user's ability to afford living expenses before recommending loans.
A new Yiddish language model, MameLoshnLM, has been developed along with an evaluation benchmark, significantly advancing AI capabilities for this minority language.
A new paper explores the challenges and advancements in continual learning, focusing on how AI models can adapt and acquire new knowledge without forgetting old information.
Research reveals that vision encoders can inadvertently identify the camera model used to capture an image, raising potential privacy and security implications.
MASS introduces a framework for multiplayer world models that maintains an authoritative shared state, crucial for developing complex and consistent AI simulations and games.
This paper proposes a paradigm where the optimizer itself acts as an agent, driving reasoning-based search across various AI artifacts like prompts, programs, and machine learning workflows.
A new research paper re-conceptualizes test-time training by breaking it down into composable modules, potentially simplifying and improving adaptive AI systems.