Netflix has acquired InterPositive, an AI startup founded by actor and director Ben Affleck, for $587 million. The company specializes in automating visual effects (VFX) and post-production workflows. This acquisition highlights Affleck's deep engagement with AI technology and reflects a growing trend among mainstream Hollywood studios to accelerate the adoption of generative AI across film and television production pipelines—particularly in high-cost areas like special effects and post-production, where AI holds significant potential to reduce costs and improve efficiency.

SenseTime has launched the SenseNova U1 Pro, a multimodal flagship model based on the NEO-unify unified architecture, which enables understanding, generation, and action to share the same Transformer backbone, eliminating separate encoder-decoder modules found in traditional multimodal models. The model supports native 8K output and designer-grade delivery capabilities, outperforming GPT Image-2 in multimodal tasks. Through Agentic Generation Loop (observe-modify-plan next step), it achieves closed-loop iterative optimization, capable of generating hundreds of error-free Chinese characters in one go and accurately simulating physical logic such as the positional relationships between Earth, Sun, and Moon. It is positioned as a delivery system for complex multimodal tasks.

Pingtouge has officially open-sourced the T-Head SAIL software stack for its Zhenwu AI chip, offering global developers full access to its five-layer architecture including drivers, compilers, high-performance libraries, and developer tools, with full compatibility across mainstream AI ecosystems. The stack emphasizes three key features: no code rewriting required, average adaptation within seven days for mainstream frameworks, and support for over 260 training/inference frameworks. The Zhenwu chip has shipped 560,000 units, serving more than 400 customers across 20 industries. Validated at scale in production environments, the software stack will be open-sourced in three phases, including core components such as communication libraries and inference engines.

A team from HKUST discovered a vulnerability in the routing mechanism of Mixture-of-Experts (MoE) models at ICML 2026. When inputs contain highly repetitive token patterns, tokens become heavily concentrated in a few experts, causing load imbalance. In 8-GPU EP deployment experiments, this black-box attack, named RepetitionCurse, increases time-to-first-token (TTFT) by 20% to 148%, worsening GPU utilization from balanced levels of 33%/34%/33% to extreme imbalances like 98%/1%/1%, turning a single GPU into a bottleneck. The attack requires no knowledge of model weights or backend deployment details, and existing defenses are largely ineffective, primarily due to the lack of load balancing constraints during inference.

NVIDIA has released DeepStream 9.1, introducing agentic AI to the field of vision AI. This version includes 13 agent skills that enable the creation of visual pipelines via natural language prompts, such as deploying multi-view 3D tracking (MV3DT). MV3DT provides globally consistent object ID tracking across cameras by projecting 2D detections into a shared 3D coordinate system. The new AutoMagicCalib (AMC) feature enables automatic camera calibration from existing videos without manual chessboard setup. The release supports Jetson Orin and Thor edge devices with JetPack 7.2 and has been unified into an open-source monorepo under CC-BY-4.0 and Apache-2.0 licenses.

NVIDIA has introduced Nemotron-3-Embed, an open-source series of embedding models, with its 8-billion-parameter version ranking first on the RTEB benchmark. Around the same time, Major League Baseball (MLB) restricted teams from using sideline iPads for AI-assisted decision-making to prevent excessive technological intervention; Meta announced it will send alerts to parents if teenagers mention self-harm or other mental health crises during AI chat sessions, strengthening online protection for minors. These developments reflect AI’s broad impact across technology development, sports, and societal ethics.

LangChain has announced the open-sourcing of an end-to-end software engineering agent factory, providing a framework and toolchain for building complete software engineering agents from scratch, accompanied by a detailed technical blog written by engineer Brace Sproul. Additionally, LangChain is promoting GPT Researcher, built on its Deep Agents framework, as the easiest entry point for creating deep research agents. These moves continue LangChain's strategic focus on open-sourcing agent development frameworks.

Youliqi has proposed UniTac, a unified cross-sensor tactile understanding and generation architecture, shifting tactile perception from post-contact feedback to pre-contact prediction. Current Vision-Language-Action (VLA) models perform poorly on deformable objects and dense contact tasks due to lack of tactile foresight, often causing damage before corrective actions can be taken. UniTac decouples tactile perception into object physical properties and sensor configurations, enabling unified cross-sensor representation. On the generation side, a Sensor-Aware DiT Projector generates post-contact tactile signals while sensors remain idle. Experiments show that VLAs integrated with UniTac can predict deformation and plan gentle interactions when grasping paper cups or towels, avoiding crushing or failed grips.

At WAIC 2026, Mech-Mind showcased its integrated 「eye-brain-hand」 embodied intelligence solution, achieving precise manipulation in industrial, retail, and home scenarios through 3D vision, the multimodal large model Mech-GPT, and the dexterous hand Mech-Hand. Its proprietary industrial-grade 3D vision can identify thin-walled and transparent objects while keeping pace with production lines; Mech-GPT accepts multimodal instructions such as natural language and images, demonstrating human-like understanding, reasoning, and generalization; Mech-Hand, centered on a world motion model, performs fine manipulations like gripping, twisting, and picking. The product has achieved large-scale deployment, with cumulative installations exceeding 27,000 units globally and serving over 100 Fortune 500 companies.

Anthropic has announced an extension of the 50% higher weekly usage limits for Claude Code until August 19, covering Pro, Max, Team, and Enterprise users. This policy is associated with the rollout of Fable 5 capacity and may become permanent. This marks Anthropic’s ongoing adjustments following the release of the Fable 5 model to alleviate capacity pressure and enhance the experience for paying users.