AI News

Hidden debt of US Big Five tech giants surged eightfold in four years to $1.65 trillion, far exceeding actual debt, mainly driven by large spending on data center leases and GPU supply contracts, e.g., Meta's hidden debt alone reaches $420 billion.....

Unitree unveils UnifoLM-OminiA-0.3, a humanoid robot AI model enabling closed-loop from multimodal interaction & perception to autonomous action. The G1 robot handles home care tasks: moving pillows, recognizing colors/counts, grasping medicine boxes, tidying clothes, storing dishes, adjusting devices.....

Google DeepMind unveils GenCeption, repurposing video generation AI for visual analysis. A single model handles depth estimation, segmentation, and 3D pose estimation in one forward pass, trained on Alibaba's Wan2.1 framework, breaking task isolation.....

Zhiyuan Robot, valued at ~$20B, is advancing its IPO with CITIC Securities as sponsor; projected 2026 revenue: RMB 4B. At WAIC 2026, it unveiled five new robots—Yuanzheng A3Ultra, Jingling G2Max, Lingxi X2EDU, Linjiedian dexterous hand, and Kutuo riding robot—embodying the "Three Intelligences in One" framework.....

OpenAI's former CTO Murati's startup launched Inkling, heavily borrowing from Chinese models like DeepSeek and Moonshot AI, yet underperforms at higher cost. This role-reversed company is valued at $12 billion, a stark contrast.....

Tencent and Chinagoods jointly launched "Chinagoods AI Navigation", the first integrated indoor-outdoor navigation solution across multiple buildings and corridors in Yiwu International Trade City. Targeting 8 million sqm and 80,000 shops, it provides AI route planning, precise indoor-outdoor navigation, and AI object recognition to find shops, solving the difficulty of locating shops.....

On July 21, Japanese AI startup Sakana AI unveiled Fugu Cyber, a network defense model achieving 86.9% on CyberGym and 72.1% on CTI-REALM benchmarks, surpassing top closed-source models like GPT-5.5-Cyber and redefining cybersecurity frontiers.....

Wojciech Gryc notes that Anthropic's Fable 5 faces competition from Chinese open-source models like Moonshot's Kimi K3 and Alibaba's Qwen 3.8, which offer comparable performance and will release open weights. While training costs involve researchers and compute, he stresses that inference costs—GPU usage per API call—are the real ongoing financial burden.....

OpenAI upgrades its youth protection measures, expanding the parent notification feature to include threats of violence and online bullying. Previously, it supported parents linking their children's accounts, setting usage limits, and alerting about self-harm risks. Now, when a minor is banned for content involving harm to others, their guardian will receive a notification. This feature was developed in collaboration with the online bullying monitoring organization Moonshot.

YouTube updated its policy to fight AI spam, classifying 'inauthentic content' into three types and fully banning monetization. The first banned type covers generic, repetitive, templated videos—often AI or computer-generated—lacking originality and simply rehashing existing content.....

Alibaba plans to launch 'Qianwen Office,' integrating QoderWork, Wukong, and MuleRun agents to target the enterprise AI office market. DingTalk's new CEO Chen Yusen, who took over in June, unified management of these products in July, shaping Alibaba's key agent-based office strategy.....

After the release of Moonlight Face's new model kimi3, the demand exceeded expectations, leading to a shortage of computing resources. New user registration is temporarily not fully open, with priority given to paid existing users. The B-end representative stated that sufficient preparations were made, but the actual growth still exceeded estimates, resulting in resource pressure. The company clearly stated that it will not sacrifice the current experience for short-term user growth.

Visual large models faced hidden barriers—SOTA required massive data and compute, reserved for top teams. On July 21, XPeng unveiled TuringViT, an efficient vision encoder that restructures architecture, data, and training, enabling reproducible, low-cost SOTA vision Transformers and democratizing advanced visual capabilities.....

Alibaba released Qwen-Image-3.0, supporting up to 4.5K-token prompts to generate complex content like formulas, geometry, logic, and multi-layer UI. It natively renders 12 languages and 20+ fonts, with 4.5x longer text vs. prior version, advancing commercial creative tasks such as film storyboards and knowledge diagrams.....

District 9 director Neill Blomkamp releases his first AI short film, 'Ghost Soldiers,' a 13-minute piece fully generated by ByteDance's Seedance 2.0. He guided creation frame by frame via text prompts, using licensed likenesses and voices of 32 real actors, with support from a professional concept artist.....

China Telecom pioneers large model for 5G wireless network planning, boosting efficiency by 50% and accuracy over 75%. Automatically generates designs and predicts results, transforming traditional manual survey methods.....

The global AI and data center expansion has triggered a surge in semiconductor demand. South Korea's exports, adjusted for working days, surged by 62.9% year-on-year in the first 20 days of July, setting a new record for the same period in history; unadjusted export volume also increased by 52.3%, while imports rose by 20%. Strong exports highlight the continued high demand for chips.

At the 2026 World AI Conference, the question of when humanoid robots will leave the exhibition hall and become practical tools was repeatedly asked. On July 19, Zhang Wei, founder of Zhuji Dynamics, countered: embodied AI is on an exponential growth curve, and robot brain systems have evolved to roughly the GPT-3 level. Zhuji Dynamics was founded in 2022.....

Microsoft is considering integrating Moonshot AI's Kimi K3 open-source model into Copilot to significantly reduce AI operational costs. As the largest open-source model globally, K3 rivals top models in coding and other benchmarks. It could replace reliance on OpenAI and Anthropic, reshaping the architecture, with internal estimates saving up to $600 million annually.....

Tencent's Hyra-1.0 agent adopts a 'light framework, broad actions' design, using recursive self-improvement to turn compute into gains. It features a Context Agent for inspiration from experience library and parallel execution Agents, targeting performance-oriented research and engineering.....