AI News — Sunday, June 21, 2026
Nobel laureate John Jumper, a key figure in AlphaFold's development, is reportedly leaving DeepMind to join rival AI research company Anthropic, signaling a significant talent shift in the industry.
OpenAI has announced its acquisition of Ona, a company focused on data collection and analysis in challenging environments, likely to enhance OpenAI's data capabilities for AI model training and deployment.
Researchers introduce DragMesh-2, a new method enabling physically plausible and dexterous hand-object interactions with articulated objects, significantly advancing robotics and virtual reality applications.
Meredith Whittaker of Signal cautions users to remember that AI chatbots are not personal friends, emphasizing the importance of privacy and critical evaluation of AI interactions.
A new service named Goose claims to offer the same functionality as Claude Code, which costs up to $200 a month, entirely for free, potentially disrupting the market for AI coding assistants.
OpenAI has introduced new Academy courses designed to equip individuals with the skills to apply AI effectively in various professional settings, catering to the evolving demands of the modern workforce.
FlowBender introduces a novel feedback-aware training mechanism that enables conditional flows to self-correct, promising more robust and adaptable generative models.
JanusMesh presents a new technique for generating 3D visual illusions rapidly and without prior examples, utilizing cross-space denoising to create compelling visual effects.
This research investigates whether World Action Models truly require complex video generation capabilities or if advanced image editing techniques suffice for their operational needs, potentially simplifying model design.
A new paper argues that current world models are missing a persistent state core, highlighting a fundamental limitation in their ability to maintain consistent understanding and memory over time.
FAPO introduces a method for fully autonomous prompt optimization within multi-step LLM pipelines, aiming to enhance the efficiency and performance of complex AI workflows.
ENPIRE proposes a framework for agentic robot policy self-improvement directly in real-world environments, paving the way for more autonomous and adaptable robotic systems.
MaineCoon is a research initiative focused on developing a real-time audio-visual social world model, aiming to create AI systems that can understand and interact with complex social environments.
This article argues that AI memory should be integrated as a fundamental product state rather than relying on prompt engineering tricks, advocating for more robust and reliable AI interactions.
New research explores the concept of 'hidden anchors' that influence decision-making in multi-agent LLM deliberation, shedding light on the underlying mechanisms of collective AI reasoning.