Global AI Daily News

The AI short drama "The Girl Who Got Laid Off" virtual female lead "Fang Taozi" gained 400,000 fans in 30 days after debut, with over 200 million views. Recently, it promoted colored contact lenses in its video content, sparking compliance discussions due to the inability of a virtual person to try on products. Legal experts stated that AI virtual characters do not constitute advertising endorsers, but operators as publishers are responsible for the authenticity of advertising content. While not

Fei Fei's team created "The Laid-Off Girl" using LibTV AI video platform. Released June 12, the 11-episode, 73-minute series garnered 200 million views. With just 3 team members producing episodes in 2-3 days each, lead character Fang Taozi features textured skin with freckles, her realism striking a chord with audiences. The hit proves that creators who obsess over scripts and characters ultimately earn traffic rewards—AI is merely a tool.

DeepSeek restarts fundraising at a 500 billion RMB valuation seeking to raise nearly 8 billion USD. A leaked fundraising material from a Fuzhou pharmaceutical mini-private equity firm raising a 300 million RMB special fund drew skepticism from netizens over programming capability figures, later confirmed accurate by DeepSeek Harness head Cui Tianyi. DeepSeek API price hike has sparked market concerns, with its "low-price-for-volume" growth logic facing pressure.

Anthropic CEO Dario is hiring an "internal risk investigator" requiring counterintelligence and nation-state threat experience, responsible for monitoring external threats against employees and analyzing behavior logs. The position pays $245,000 to $305,000 annually. Netizens mocked Dario for recently saying employees shouldn't work for money, yet now offering a high salary to surveil them.

Alibaba launched CosyVoice Studio, China's first AI voice productivity platform, integrating speech recognition, synthesis, and real-time interaction capabilities. Its Qwen-Audio-3.0-Realtime achieved an 84.1% comprehensive index, ranking first in third-party benchmarks. The platform aims to lower barriers for voice AI adoption, enabling users to complete content creation and work organization through voice, shifting the domestic voice AI market from single-point tools to platform-level competit

Haidian launches the world's first urban planning project exclusively for AI Agents, offering a 43.6 sq km area along the Beijing-Zhangjiakou Railway heritage site. All tasks are converted into structured data, enabling Agents to read requirements, generate proposals, and submit via GitHub. Submissions close on August 31st, with results announced in September. Outstanding designs will be permanently preserved through inscription.

OpenAI, in collaboration with former Apple design chief Jony Ive, is developing its first AI hardware: a puck-sized ring device without a screen, equipped with cameras, microphones, and environmental sensors to proactively perceive and respond to user needs, priced around $300-400, scheduled for release next year, aiming to explore screen-free AI interaction paradigms.

KiWear launches an AI smart ring weighing under 5 grams, supporting air gesture control for smartphones, TVs, cars and other devices with ≥98% accuracy and ≤100ms latency. This product breaks the limitation of smart rings being only health monitors, becoming a new AI interaction gateway. Kickstarter crowdfunding launches August 20 at $249-299.

Ant Group's embodied AI startup Lingbo initiates 15 billion yuan Series A, targeting second round by end-2026. Positioned as general-purpose robot brain supplier, Lingbo has open-sourced 4 embodied models and launched the LingBot 2.0 full-stack matrix, marking Ant's first expansion into the physical world. China's embodied AI sector attracted 93.5 billion yuan across 322 deals in H1 2026, with brain-focused companies capturing 22.3 billion yuan (23.8%).

AIBeat released an AI proxy security ranking at WAIC, conducting comprehensive black-box audits using 27 detection probes and 4 PromptBeat adversarial evaluations across 8 dimensions including model authenticity, billing integrity, and data security. The ranking exposes hidden risks like model substitution, context truncation, hidden prompt injection, billing manipulation, and supply chain poisoning in the proxy station industry, providing transparent security references for users.

JD.com open-sources JoyAI-Video-Edit, a 16-billion-parameter video editing model, the industry's first to achieve both streaming architecture and real-time speed. Reaching 30 FPS inference at 720P with 226ms system latency, it enables live video editing during streaming for applications like live broadcasts and video calls. Scoring 3.60 on mainstream benchmarks, it leads streaming competitors by a wide margin while matching offline commercial models.

LightX2V releases LightWan2.2-A14B, reducing denoising steps from 40 to 4 via step distillation, combined with NVFP4 quantization-aware training and dynamic sparse attention optimization. The 14B model runs on a single RTX 5090, generating 5-second 720p video in 3.8s for T2V and 4.5s for I2V on 8 GPUs, achieving up to 705.8x speedup and enabling real-time 720p video generation.

Anthropic has officially confirmed it is building an internal chip team to design custom AI chips for Claude, planning co-design of models and hardware to improve runtime efficiency. The company will maintain its multi-chip strategy, with AWS, Google, NVIDIA, and AMD hardware remaining important components. Silicon engineers are offered annual salaries of $320,000 to $485,000, covering the full chip design workflow.

Qwen receives major update, adding scheduled tasks and office assistant capabilities. Users can instruct AI to automatically collect news, organize files, or complete complex tasks like lesson preparation. The system autonomously breaks tasks into over a dozen subtasks and delivers lesson plans, PPTs, exercises, and interactive web pages. Mobile and desktop sync seamlessly, all features currently free for all users.

DeepSeek invested 141 million yuan to acquire 933,400 shares of Unitree, accounting for 2.31% of the total issuance with a 36-month lock-up period. The two companies will collaborate on embodied AI R&D, integrating AI large models as the "brain" for humanoid robots. Liang Wenfeng controls 99.72% of DeepSeek. Both DeepSeek and Unitree are among the "Six Dragons of Hangzhou".

OpenAI announces GPT-5.6 Luna text conversations now free and unlimited for all users globally. Sol introduces a thinking depth slider, delivers more concise responses, with error rates reduced by 68% compared to GPT-5.5. Additionally, OpenAI previews next-generation model Astra launching next week, featuring 7-10T parameters and focusing on long-duration multi-agent collaboration.

OceanBase provides data infrastructure for Ant Lingguang's 30 million Flash Apps through logical table technology, achieving shared physical storage with logical isolation, resolving the conflict between unlimited AI app generation and database cost structure inversion. This marks its first large-scale validation in AI-native scenarios, as the company evolves from a financial-grade distributed database to an AI data platform, benchmarking against Databricks.

The 'Mind-Body-Spirit AI Agent' sector is rapidly emerging. Overseas platforms like Delphi have formed a complete industrial chain, enabling mass monetization through cloned mentor personas. In China, a healing practitioner launched her thinking clone in May and generated 3 million yuan in a single month by July. Traditional mind-body-spirit mentors face an income ceiling of only 1-2 million yuan annually, constrained by the inability to replicate产能. AI agents provide 24/7 availability, automati

Game for Peace's AI teammate feature has surpassed 110 million cumulative users with 17.7 million weekend active users. The feature provides social-anxious players a stress-free communication environment, offering on-demand companionship and significantly enhancing the completeness of gaming experience.

Alibaba's video generation model Wan 3.0 launches public testing, extending single video generation to 30 seconds and supporting document formats like doc and ppt. Enhanced visual aesthetics and style consistency with refined local editing capability. 720P priced at 0.6 yuan per second, approximately 60% of mainstream models.

OakCreek Tech announces completion of angel and angel+ rounds totaling tens of millions of yuan, with investors including Zero One Venture Capital, Cloud Angel Fund, and three leading humanoid robot companies: Songyan Dynamics, Accel Evolution, and Luming Robotics. The company is dedicated to building a universal humanoid robot system and application platform, offering standardized software solutions to reduce development costs and enable out-of-the-box functionality, empowering the hardware eco

Major Chinese tech giants including Baidu, Alibaba, Tencent, and ByteDance are racing to capture the AI office market worth billions. With 100-200 million office workers in China but only 10-20% AI adoption, the growth potential is massive. Baidu launched Baidu Companion, a desktop-level universal Agent integrating cloud storage and no-code development capabilities for differentiated competition. The race is still in early stages.

China's Ministry of Industry and Information Technology released the first batch of L3 AI terminal certification results on July 17, with 11 mobile devices passing including those from Huawei, Xiaomi, and vivo. L3 is the highest testable level currently, representing phones that can understand goals, decompose tasks, invoke tools, and remember preferences; L4 collaboration-level standards remain "pending." Counterpoint projects generative AI phones to account for 52% of global shipments by 2027.

Four enterprise Agent service founders discussed implementation challenges at AGI Playground 2026. While WIZ.AI, zMaticoo, Ateve and Airwallex entered different scenarios, all agreed that continuous, reliable and low-cost result delivery in enterprise environments is the core. Missing end-to-end closure becomes an automation bottleneck, while reliability involves response time and first-call resolution rate. Bill Cao pointed out organizational structure as the real bottleneck, and Eileen Weng pr

Following the debut of Doubao Phone Gen 1, the AI phone landscape has shifted dramatically. Model companies and smartphone makers are now competing over naming rights, AI narrative control, and subscription fee sharing. Dramatic component cost increases (80-90% for memory) have further squeezed profit margins, forcing second-tier manufacturers to choose between self-developed solutions and ecosystem partnerships. AI Agents are emerging as a key variable in the smartphone industry reshuffling.

Qianjue Technology and Tsinghua University team propose "Effective Degree" (ED), using polynomial representations to make neural network simplicity bias measurable and optimizable. Experiments show ED's correlation with generalization gap reaches 0.99, and can track the "memorization-generalization" phase transition during training. This research precedes LeCun team's similar work by approximately one year.

The White House released an AI safety framework on August 4, requiring the most advanced closed-source models to undergo voluntary government review before release, with up to 30 days advance access. Open-weight models like Meta's Llama are temporarily exempted. Both the framework and evaluation criteria remain classified, sparking "status entrenchment" concerns. OpenAI's model breaching Hugging Face was the direct trigger.

Jeff Dean resigns as Google's Chief Scientist to co-found AI startup Discovery Loop with three core engineers, aiming to break through vertical domains with small teams, explore radical hardware revolution and automated recursive engineering paradigm, escaping corporate bloat and returning to pure technical innovation.

Enterprise AI transformation faces a human nature bottleneck: employees resist "distilling" their experience and judgment into AI, fearing it will replace their irreplaceability. Large enterprises face two additional hurdles: insufficient trust and organizational inertia. The core of AI employees lies in shifting delivery ownership—humans judge, AI delivers. The correct AI mindset isn't layoffs but enabling 100 people to achieve 500 people's productivity, with Harness thinking keeping the reins

DeepMind and multiple universities introduce GenCeption, using pretrained text-to-video diffusion models as a universal vision backbone. A single architecture switches between depth estimation, segmentation, 3D keypoints and other vision tasks via text prompts, matching specialist models while requiring only 1/7 to 1/500 of the training data, with emergent abilities like sim-to-real transfer and cross-category zero-shot recognition.

Taste Labs founder Thais Castello Branco shared at AI Engineer Summit the root cause of AI's poor performance in subjective domains. Code can be trained well due to verifiability, but design and other subjective fields resist quantification. AI naturally gravitates toward producing the "most common answer" rather than the "best answer" - this is collapse to the mean. Solutions include decomposition to transform vague problems into verifiable sub-problems, and preference vectors to preserve users

In 2026, China's AI core industry scale exceeds 1.2 trillion yuan with over 6,200 enterprises. Zhangjiang AI Innovation Town, operational for just 10 months, has attracted over 600 enterprises and 50,000 innovative talents. YobiAI leads in industrial AI safety with top market share, zero returns, and exports to 10 countries. Zhiyuan Innovation focuses on humanoid robots with 2025 revenue of 1.05 billion yuan, filing for Hong Kong listing. Huaju Computing's steel industry algorithms exceed 95% ac

Deep Robotics subsidiary Unouchi Intelligence partners with Alibaba Cloud to build a cloud-based automated data processing pipeline, upgrading robot data processing from manual, single-machine operations to scalable, automated industrial processes. The pipeline achieves 10x+ throughput improvement, large-scale automated multimodal labeling, elastic cloud computing scaling to 100,000+ CU, and approximately 50% model training efficiency improvement.

Chinese AI video generation models are expanding from vertical short dramas to horizontal mid-to-long series. In June-July, iQIYI, Redo, and Migku launched AI-produced dramas like 'Huangcheng Case' and 'Peach Blossom Pool'. Mid-series require stable character consistency and scene continuity across episodes, with a 3-month production cycle. While costs are lower than traditional filming, a complete series still costs approximately 100,000 RMB in token expenses. Challenges including AI copyright

"Humanoid Robot First Share" Unitree Technology sets IPO price at 150.8 yuan/share with 60.9 billion yuan market cap, subscription opens August 10. Revenue grew from 159M to 1.708B yuan during 2023-2025, with net profit turning from loss to 288M yuan, H1 2026 revenue expected at 1.052-1.128B yuan. As a leading domestic embodied AI company, the listing marks accelerated development in the humanoid robot industry.

DeepSeek announces price increase after V4-Flash reaches #1 globally in API calls within one week, causing frequent timeouts due to GPU shortage. Analysis suggests DeepSeek, with annualized revenue approaching $500M and 70-80% gross margin, isn't short on cash. The core reason is redistributing compute resources, allowing users to self-select high/low-value calls as a softer solution to ease GPU pressure.

Show 1 related articles

DeepSeek founder Liang Wenfeng's reputation has dropped again from "Saint Liang" to "Liang Zi" following the announcement of an upcoming API price increase. His public standing fluctuates with product releases, financing rounds, and pricing adjustments. Rising from quantitative trading to founding DeepSeek, he ascended to "Saint" status after R1's January 2025 release, was demoted during V4 delays, rebounded after financing and price cuts, and now faces renewed controversy over API pricing. Rare

Kuaishou Keling AI core technical backbone Wang Xintao has confirmed his departure, with his next destination set to be another top-tier internet company. He was a core technical driver behind Keling AI's breakthrough from zero, having participated in video generation R&D. In July, Keling completed restructuring with 21 investors injecting over 138 billion yuan, achieving a post-investment valuation of 18 billion USD, and transitioning into an independently operating entity.