Overview
Today’s AI landscape is a study in contrasts—creativity meets labor strife, and research breakthroughs land alongside infrastructure tweaks. Google DeepMind made headlines on two fronts: a headline-grabbing partnership with indie film powerhouse A24, and a rocky start to unionization talks that left employees frustrated. Meanwhile, the open-source and academic corners of AI delivered notable advances in robotics and computer vision that could reshape how autonomous systems learn and track people in the wild.
The A24 deal signals a growing appetite among AI labs to collaborate with creative industries—an area where tools like generative video and script analysis are beginning to move beyond hype. But the unionization friction at DeepMind serves as a reminder that as AI expands, the humans building it are demanding a seat at the table. On the research side, a new unified training framework for vision-language-action models (VLAFlow) promises to cut through the noise of disparate robot-learning setups, while a multi-person 3D mesh tracking method (Multi-THuMBS) tackles the persistent challenge of tracking people across video cuts and occlusions. For anyone tracking the tools behind these advances, platforms like GetAI Business help surface the emerging models and techniques shaping the field.
Today's Big News
Google DeepMind Partners With A24 for First-of-Its-Kind AI Research Collaboration
The deal between DeepMind and A24—the studio behind Everything Everywhere All at Once and Moonlight—marks one of the most high-profile crossovers between AI research and independent cinema. While specific project details remain under wraps, the partnership is expected to explore how generative models can assist in storytelling, visual effects, and pre-production processes. This move reflects a broader trend: AI labs are increasingly seeking real-world creative use cases beyond text and image generation.
DeepMind Unionization Negotiations Hit Early Roadblocks
According to a WIRED report, Wednesday’s bargaining session between DeepMind leadership and employees pushing for unionization revealed a widening trust gap. Workers voiced frustration that executives appeared unwilling to engage substantively with their demands for better pay transparency, job security, and input into how AI models are deployed. The rocky start underscores a growing tension in the tech industry: as AI capabilities accelerate, the human workforce behind them is organizing for more influence.
VLAFlow: A Unified Training Framework for Vision-Language-Action Models
A new paper on arXiv (2607.01586) introduces VLAFlow, a co-training framework designed to harmonize data and architectures across robot-learning paradigms. By aligning future latent representations and combining multiple pre-training datasets, VLAFlow achieves stronger generalization on manipulation tasks than models trained on single data sources. This could be a key step toward more scalable and transferable robotic skills, especially for service and industrial robots that must adapt to varied environments.
Multi-THuMBS: Tracking 3D Human Meshes Across Multiple Video Shots
Multi-person 3D human mesh tracking has long been stymied by occlusions and scene changes, but the Multi-THuMBS method (arXiv:2607.01626) proposes a new approach that works across video shot boundaries. It leverages temporal coherence and multi-view reasoning to maintain identity and pose continuity even when people leave the frame and reappear. This advance has practical implications for sports analytics, surveillance, and virtual production—any domain that needs consistent 3D human tracking in unconstrained footage.