Global AI Daily News

Since 2026, AI companies have been mass purchasing secondhand printed books through intermediaries like ISBNdb, digitizing them via scanning, then destroying the originals. Anthropic's "Project Panama" was exposed: legally purchased books are rapidly scanned and then processed as recycled materials. U.S. courts ruled this constitutes "transformative fair use." Ironicallly, scanning and destroying purchased books is legal, while using pirated copies constitutes infringement. Humanity has publishe

ByteDance completed a major ToB restructuring, merging Feishu's product team into Doubao and its commercialization team into Volcano Engine. This marks the most fundamental mission shift in Feishu's history—from redefining digital collaboration to redefining what advanced enterprises need in the AI era. Feishu's accumulated organizational data serves as the foundation for agents entering enterprises, and with no mature methodology in the market, this restructuring offers Feishu an opportunity to

Tiiny AI secures tens of millions of dollars in angel funding from top-tier investors including Shunwei Capital and Matrix Partners. The company has partnered with a leading global PC manufacturer for OS-level deep collaboration, securing a 10,000+ unit pre-installation order with per-unit license fees exceeding those of a well-known PC operating system. Its 300-gram Tiiny AI Pocket raised over $3 million on Kickstarter, running billion-parameter models on just 10W power consumption, marking per

A blogger worked while hiking for 4 hours in the redwood forest of Northern California, using ChatGPT's voice mode through AirPods. He reported completing more work in those 4 hours than in 8 hours at the office. Sam Altman reshared and endorsed this new way of working, noting that voice interaction is blurring the boundaries between work and life.

Muka Robotics releases LJM world model, trained on just 32 Zhenwu 810E GPUs, ranking 2nd globally on WorldArena under anonymous code SisyphusWorld, with all four LIBERO metrics achieving new SOTA. The model employs a dual-expert architecture that first reasons about interactions in latent space before generating video, solving the core issue of video diffusion models prioritizing visuals over physics, enabling robots to truly understand the relationship between actions and physical outcomes.

Show 1 related articles

Chi Xiaowei, a 95-born HKUST PhD, founded MUKA Robotics in April, focusing on world model research with backing from BlueRun Ventures and ZZDL. The company develops a universal robot intelligence system with visual generalization models as the core, enabling robots to transition from closed scenarios to open environments. Chi is the first author of the open-sourced WoW embodied world model released in October 2025.

Opus 5's 'Challenge Loop' game prompt went viral on 𝕏, attracting millions of viewers. This prompt employs a three-layer architecture: main agent decomposes tasks, sub-agents handle creation, and a judge agent verifies results. User Anshu recreated a space exploration game in just 24 hours, with all textures, animations, and sounds generated by AI without external materials.

Microsoft reports Q4 FY2026 earnings with revenue of $90B up 18% YoY and net income of $35.8B up 31%. Azure revenue grows 43% to exceed $100B annually for the first time, while Copilot paid seats surpass 30 million. Capital expenditures reached $41B, up 70% YoY, with $3.2B in Anthropic investment gains. After-hours shares rose 8.83%.

The US government has ordered contractors to eliminate all Anthropic products by August 31, 2026, including Claude web apps, APIs, and Claude Code. Silicon Valley companies are widely dissatisfied with Anthropic due to its competitive product launches, closed ecosystem advocacy, and data retention policy changes, with many shifting to alternatives like OpenAI.

GPT-5.6 Sol autonomously rewrites OpenAI's production GPU kernel, reducing service costs by 20% and improving token generation efficiency by over 15%, with optimizations targeting the live production system serving billions of daily user requests.

VAST delivered a keynote at SIGGRAPH 2026, with five papers accepted to Technical Papers and winning the Best in Show award at Real-Time Live!. Chief Scientist Dr. Cao Yanpei was named to MIT Technology Review's TR35 China list. Flagship model Tripo P1.0 generates clean-topology 3D assets ready for game engines in 2 seconds, serving over 20 million global users.

SenseTime releases Seko 3.0 AI video creation platform, achieving 1 million users (26x growth), 1,500 enterprise clients, and 89x revenue increase in one year. The platform enables integrated production from script to multilingual distribution via Canvas, Agent, and Skill modules, allowing creators to generate animated series and expand globally. The OPC base offers zero-cost entrepreneurship support and reusable skill templates, lowering creative barriers. China's micro-short drama market is pr

SemVID is a training-free token pruning method for video temporal grounding that classifies visual tokens into Object, Motion, and Context roles. By preserving only 12.5% of tokens, it maintains evidence chain integrity and solves temporal fragmentation issues under high compression, accepted by ECCV 2026.

Infinigence releases PDD cross-cluster heterogeneous inference architecture, innovatively adopting P-RLD-MD three-tier separation, masking KV transmission latency through relay mechanism, achieving 51.5% reduction in first token latency and 37.5% reduction in per-token cost.

Researchers from NTU and CUHK released Apple-\u03c0, a physics reasoning benchmark for video models, featuring the Orchard dataset with 400 cases across 5 evaluation tracks. Testing 11 models with 6,000 responses revealed that even the best video model Seedance 2.0 scored only 0.473, while the strongest unified model achieved approximately 0.40 in reasoning tasks, indicating that consistently following physical laws remains a core bottleneck.

WorkBuddy launches dual human-AI writing feature, integrating AI-powered Office within the app. AI understands ongoing document changes, supports precise selection-based operations, and collaborates directly in PPT, Word, and Excel, reducing copy-paste tasks. It also introduces an Agent-exclusive mailbox, enabling AI to proactively engage in business workflows.

Top universities worldwide have jointly developed JarvisHub, an open-source workbench for long-horizon multimodal creation. It represents creative projects as editable multimodal asset graphs, uses protocol bridging to control agent operations, enabling tool orchestration, version management and local restoration, allowing creators to continuously advance complex projects like images, videos and websites.

SemiAnalysis founder Dylan Patel presents seven core judgments on AI infrastructure: AI competition is shifting from model capability to cost and delivery capability. Cross-layer collaboration can yield 100x efficiency gains; cheaper tokens drive higher total compute demand; memory faces structural shortages while timely power availability becomes critical; Nvidia needs a multi-polar customer structure. Anthropic and OpenAI's combined compute demand may exceed 100GW by 2030, versus roughly 6GW c

OpenAI open-sources Codex Security CLI, an AI tool that automatically scans code for vulnerabilities and provides fix recommendations. Originally a closed-source security plugin in Codex, it's now an independent tool for teams and enterprises, supporting CI/CD integration and requiring Node.js 22+ and Python 3.10+.

A 10-person team at Chance AI developed Chance Vision 1.5, achieving 86.9% on the MMMU-Pro benchmark and surpassing GPT, Gemini, and high-level human experts. Their VIC technology enables AI to continuously revisit original images to verify evidence, solving "visual grounding drift" and advancing from "seeing" to "insight." The company serves approximately 300,000 global users.

OpenAI hardware roadmap revealed: first screenless speaker launching early 2027 at \$200-300, featuring camera and mechanical structure; phone mass production moved up to H1 2027 with custom MediaTek chip, ISP as top specification. Supply chain includes Luxshare Precision and Sunny Optical, with analyst predicting 30 million units combined for 2027-2028.

OpenAI releases GPT-5.6 Sol self-optimization results: inference service deployment costs reduced by 20%, token generation efficiency improved by over 15%. Former executive Lillian Weng rumored to return to assist AI self-evolution research. President Greg Brockman hints at AI hardware in development. OpenAI surpasses 1 billion users, serving 2 million enterprises.

Show 1 related articles

After leaving Thinking Machines Lab due to health reasons, Lilian Weng has quickly returned to OpenAI to lead the RSI (Recursive Self-Improvement) team. The project aims to enable models to train and improve their successor models, representing a top-priority research direction at OpenAI. With Weng's return, three of TML's six co-founders have now rejoined OpenAI.

Agnes AI launches domestic website at agnes-ai.cn with API service adjustments to streamline access for Chinese users. The Agnes 2.5 Pro Alpha ranks 9th among 153 models on Artificial Analysis comprehensive capability leaderboard with an Intelligence Index of 39, supporting 1M context window. The Flash version remains free, with weekly multimodal token processing reaching 8.16 trillion.

Zhiyuan's self-developed omni-modal large model WITA-Omni Preview tops the DailyOmni global leaderboard with a score of 85.21, surpassing major competitors including Qwen, Gemini, and NVIDIA. The model achieves simultaneous interaction capabilities of "thinking while listening, speaking while moving," eliminating the stiff pattern of "thinking first, then speaking, then moving," establishing leadership in speech interaction and embodied intelligence.

Moonshot releases open-source 2.8T parameter model K3, featuring KDA attention (3:1 hybrid) and AttnRes residual connections, reducing KV cache by 75% and achieving 6x faster decoding at 1M token context with 25% training efficiency improvement. API pricing is $3/$15 per million tokens input/output, requiring 64+ accelerator cards for deployment. The project cost $19 across 26 iterations to create an interactive Naruto Rasengan webpage.

Show 1 related articles

Kimi releases K3 large language model with 2.78 trillion total parameters, 104.2 billion activated parameters, 896 routing experts, supporting 1M token context. Through three architectural innovations—KDA linear attention, AttnRes residual optimization, and LatentMoE expert compression—K3 solves challenges in 3T-level model training, achieving approximately 2.5x Scaling Efficiency improvement, demonstrating Chinese labs' capability to redesign frontier models.

Sangfor AI, the AI security agent from Sangfor, completed 1,301 out of 1,507 vulnerability challenges in CyberGym evaluation with an 86.3% success rate, ranking fourth globally and first domestically. Employing the Agent Swarm architecture with evidence governance mechanism, it achieved 87.2% and 77.7% success rates in ARVO and OSS-Fuzz scenarios respectively, surpassing OpenAI and Anthropic using entirely domestic technology stack.

Jia Yangqing founded Intent Lab launching Fleet, a software engineering AI team, achieving 6.3x performance improvement in TensorRT-LLM optimization; Andrew Ng created LearnVector with $100M Coursera investment, focusing on personalized AI education; Sonia Joseph established World Mechanics building interpretable foundation models for physical AI. All three announced startups on the same day, covering software engineering, education, and physical world domains.

Li Kaifu's 01.AI has pivoted four times in three years, from pursuing GPT-5 to abandoning frontier models, handing over pre-training team to Alibaba, and now targeting Palantir. With revenue exceeding 100 million RMB and B-end accounting for 70%, the frequent strategic shifts have drawn industry scrutiny.

Tencent Video is internally testing WorkSolo, an AIGC creation platform featuring AI short dramas, interactive gaming, and free canvas modes, operating on a membership subscription plus credit-based computing model. The platform differentiates from WorkRally by focusing on lightweight short drama and interactive content rapid production, while WorkRally targets premium manhua drama industrialization. Creators can distribute finished works through Tencent Video channels and participate in revenue

Anthropic releases research showing Claude Mythos Preview, using 60 hours and ~$100k API cost, discovered a nontrivial automorphism vulnerability in HAWK post-quantum signature algorithm, reducing key strength from 2^64 to 2^38 (44 trillion times faster); also invented "Möbius Bridge" attack on simplified AES-128, achieving 200-800x speedup.

AI blogger Karl's AI Workshop releases partner-skill 2.0, an agent collaboration tool integrating three roles (deep_reasoner, fast_worker, arbiter) with cross-project model mixing for Claude Code and Codex, automatically routing tasks to optimal models to cut costs. It features local config UI, Goal hash verification, and evidence chain receipts. The author reduced weekly API costs from $500 to $400, achieving 50% cost reduction in the second week.

SpaceXAI releases Grok Voice Think Fast 2.0 voice model, ranking first in voice agent benchmark with 56.5% score and 82.9% overall score. Response time is only 0.70 seconds, fastest among top five. Priced at $0.08 per minute, approximately 45% of GPT-Realtime-2.1 High. Default model upgrade scheduled for August 5.

Xiaohongshu and Peking University propose UltraEP, achieving real-time MoE expert load balancing within 300 microseconds, with training/inference throughput reaching 94.3% of ideal performance, 1.49x improvement over Megatron-LM/SGLang.

NanoWork launches enterprise AI agent solution enabling multi-agent collaboration for complex tasks like competitive analysis, content creation and web development. The platform integrates mainstream frameworks including Claude Code and Codex, with 1,000+ models auto-scheduled and 500+ job experts ready to use, upgrading AI from copilot to enterprise infrastructure.

Ant Group's Alipay Health AI assistant "A Fu" has upgraded its "AI Food Photo" feature, allowing users to take photos to automatically identify food items, estimate calories and nutrition, get dietary recommendations, and save data to their personal health records as a "calorie account." Combined with body fat scales and AI exercise guidance, this forms a comprehensive scientific weight loss service loop. Currently, nearly 1.5 million people nationwide have joined the weight loss initiative, los

A joint team from Tsinghua University, Shanghai Qi Zhi Institute, and Cambridge University proposes the LoC Behavioural Framework, decomposing AI loss of control into three conditions: misaligned motive, harm-enabling capability, and monitoring evasion. Tested across 13 dimensions with 29 tasks, the framework achieves a 0.84 correlation coefficient between LoC scores and long-horizon scenario failure rates, enabling early risk pathway identification. Vulnerability-focused fine-tuning combined wi

vivo BlueImage Lab introduces SmartPhotoCrafter, a framework leveraging three-step training combined with reinforcement learning for automated photographic image optimization. It comprises Image Critic and Photographic Artist modules that autonomously analyze image quality, diagnose issues, and generate optimized results without human intervention, covering both image restoration and color grading tasks.

Encore AI raises $30M Series A led by Team8. The company develops "interaction mining" technology that trains AI voice agents by analyzing corporate customer dialogue data, enabling independent or collaborative customer service. It currently serves 40+ financial clients with ARR growing over 5x, planning to expand US sales operations.

Google DeepMind has dissolved its AlphaFold team, redistributing staff to Gemini and other projects. Nobel laureate John Jumper and other key researchers have joined Anthropic. AlphaFold predicted over 200 million protein structures and was cited in over 35,000 papers. This move signals a strategic shift toward Gemini and AI agents.

Wapoo, an AI video co-creation social platform founded by former TikTok product manager Yang Mingyu, has secured nearly 10 million USD in angel funding from a strategic investor under an internet group. Launched in May in the US, Wapoo enables users to collaboratively rewrite the same video with friends, hiding generative AI behind simple operations and transforming video publishing from an endpoint into a starting point. Targeting Gen Alpha users, the platform focuses on lightweight co-creation

Moonshot AI releases open-source K3 large language model with 2.8 trillion parameters, ranking third globally in benchmarks. While weights are freely downloadable for research and most commercial use, actual deployment requires 64+ accelerator cards costing over $2.4 million with 45kW power consumption, far beyond typical users. Enterprises benefit more from API access at thousands of dollars monthly, versus millions in hardware investment.