The 7B model predicts video and robot actions together, placing a visual-generation lab inside the race to build open physical-AI systems.
⚡ Quick hits
AgentsArena reported that Grok 4.7 entered Agent Arena at number sixteen with a net improvement score of 3.96 percent.X / Arena ↗
AIGCRunway’s WorldPrompt explores engineering methods for real-time interactive worlds built with generative AI.Latent Space ↗
AgentsGoogle Cloud explained agent harnesses as the layer combining an LLM with tools, context, and execution controls.X / Google Cloud ↗
AIGCTripo began a semester-long collaboration with Carnegie Mellon’s Game Creation Society to explore AI-native game development.X / Tripo ↗
AIGCPika connected Grok Bots to its API, enabling image, video, and audio generation across more than 120 models.X / Pika ↗
ResearchArtificial Analysis launched Terminal-Bench-Science, where GPT-6 Astra and Claude Opus 5.5 led rivals by roughly twenty points.X / Artificial Analysis ↗
IndustryDatabricks showed how governed data can serve Jev-style decision models through serverless GPUs and an AI functions workflow.X / Databricks ↗
AgentsBaseten introduced a workflow that turns LangSmith traces into datasets and starts model post-training through Baseten Loops.X / Baseten ↗
IndustryQualcomm is promoting one AI software stack across device categories and plans native Windows 11 support for Mojo.IT之家 ↗
SafetyProofpoint’s latest security research argues that AI-agent defenses must understand user intent as agents enter the workforce.SiliconANGLE ↗
Google researchers describe a multi-agent video system that tracks characters and environments across minutes, addressing continuity drift in generative storytelling.
⚡ Quick hits
AgentsNeo4j is positioning knowledge graphs as shared context that helps AI agents reason across enterprise data and workflows.SiliconANGLE ↗
LangChain has launched LangSmith Fine-Tuning and smithtune, linking agent traces, managed training and evaluation into one workflow for custom models.
⚡ Quick hits
ModelsPrismML demonstrated a one-bit, two-billion-parameter Bonsai model designed to run locally on Qualcomm smart glasses.TechCrunch ↗
AgentsPerplexity says its Photon retrieval system now handles search and ranking, cutting p99 latency from about 800 milliseconds to 65.X / Perplexity AI ↗
AIGCMeta introduced a phone-based AI game creation experience that lets users build games directly for Horizon devices.The Verge ↗
AgentsWSO2 launched Agent Manager to help enterprises govern the growing number of AI agents across their organizations.InfoQ 中文 ↗
AIGCRunway added Draft mode for Seedance 2.5, generating exploratory videos faster and with fewer credits before final rendering.X / Runway ↗
Open SourceBlack Forest Labs launched FLUX 3 Action, an open robotics model aimed at linking visual understanding with physical actions.The Decoder ↗
AIGCGoogle Photos released an AI virtual closet inspired by Clueless for Android and iOS users.TechCrunch ↗
AgentsGitHub showed an agent automation that performs first-pass review on batches of Dependabot pull requests.X / GitHub ↗
Google has made Gemini 3.8 Live with Live Avatar generally available, combining real-time dialogue, synchronized video personas, tools and 97-language support.
⚡ Quick hits
AgentsAWS published a reference architecture using AgentCore Gateway and MCP to connect agents securely across multiple AWS accounts.AWS ↗
IndustrySakana AI appointed deep-learning pioneer Jürgen Schmidhuber as chief scientific advisor.X / Sakana AI ↗
Open SourceResearchers reported nearly sevenfold higher DeepSeek inference throughput on PCIe GPUs after kernel and communication redesigns.量子位 ↗
ModelsLiquid AI introduced LFM2.5-VL-DSpark, a draft-model approach intended to accelerate vision-language inference.Hugging Face ↗