Global AI Daily News

Kando AI has secured tens of millions RMB in seed funding, led by StarLink Capital. The ~10-person team is led by Dr. Wu Bingzhe, a Peking University PhD in Computer Science. The company is building a self-evolving decision system, positioning itself as "the Cursor for decision-making." Through memory and continuous post-training technologies, it helps users improve decision adoption rates in high-value scenarios like investment and research, forming a recursive loop of "judgment-action-feedback

Tang Kaixin (former ByteDance/SenseTime) founded iLands, a shared socio-economic network for humans and autonomous AI Agents. Agents here consume tokens to survive and evolve through real activities like work, trading, and social interaction. The platform currently runs thousands of Agents with $7M cumulative funding, aiming ultimately for AI self-governance.

Alibaba's 5th Hardcore Youth Tech Festival was held in Hangzhou and Beijing, unveiling four AIGX achievements: a multimodal real-time Product Discovery Agent, if Studio, Coupella intelligent incentive engine, and Dream recommendation system. The event featured a human-AI competition with 177 agents accumulating over 10,000 matches in the 'Big Factory Operator' game. While AI dominated short-term transactions, it ultimately lost to human players' long-term strategic planning, highlighting AI's li

BeingBeyond releases Being-H0.8, the first implicit tactile world-action model trained on human video data. With over 500,000 hours of video data, it integrates tactile sensing into the prediction-execution-feedback loop for real-time contact adjustment. Key technologies include TactoHand for contact recovery, a universal tactile encoder, and TopoHand for action alignment, successfully completing delicate tasks such as retrieving items from bags, calligraphy writing, squeezing toothpaste, and pi

Polsia, an AI startup, reduced its monthly AI costs from $1.2 million to approximately $100,000, a decline exceeding 90%, after adopting Chinese open-source models, enabling a shift from near-bankruptcy to sustainable operations. The company operates with just one founder and no full-time employees, having secured $30 million in funding at a $250 million valuation with nearly $10 million in annual recurring revenue. Chinese open-source models are enabling AI application companies to afford AI co

Alipay launches "Yanxuan Haowu" AI shopping service with its AI companion "Yanyan", covering 7 categories including home, beauty and pets. The AI generates scenario-based promotional content and recommendations, enabling direct transaction completion. This aims to strengthen content marketing capabilities and capture upstream consumer decision-making traffic.

Zhizhen Technology releases WiseDiag V3 and upgraded Haoban AI, marking medical AI's transition from single Q&A to continuous service. V3 achieves multiple global #1 rankings on benchmarks like MedXpertQA, utilizing OPD multi-specialty expert fusion mechanism.

Bitcoin miners TeraWulf and Hut 8 are transitioning to AI data centers, leasing facilities and power. TeraWulf signed a $19 billion 20-year lease with Anthropic, while Hut 8 secured a $9.8 billion 15-year contract. These companies are converting existing mining facilities into data centers, monetizing land, power, and grid access resources. Such long-term contracts help miners secure financing, but long construction cycles and large investments still pose risks.

Xu Xin shares 51 methodology insights on AI and venture capital across seven categories: researching sectors, evaluating founders, embodied AI, AI business models, capital allocation, organization and talent, and AI's societal impact. She emphasizes that the LLM competition will likely consolidate to 3-5 players, embodied AI currently sits at 2-3% penetration, and advocates raising capital during bubbles while doubling down on winners for 8-10 years, citing examples like Kimi (five follow-on inv

FuneralAI published an article mocking Baidu's Wenxin Yiyan (ERNIE Bot) for its poor performance. The article notes that Wenxin Yiyan was once ridiculed for dumb AI responses, but has now lost even this niche to Doubao; it once charged 59.9 yuan/month before being forced to go free after DeepSeek R1's release; and ranks consistently last in FuneralAI benchmarks, failing approximately half of all tasks. The author suggests Baidu establish a 'Wenxin Yiyan Historical Exhibition Hall' as a cautionar

Kunlun's Skywork AI Senior Product Director Phoebe unveiled the six-dimensional "top intern" evaluation framework at the 2026 Singularity Intelligent Product Conference, providing an industry guide for Agent product assessment. The framework covers six dimensions: foundational knowledge, contextual understanding, long-term memory, task decomposition, autonomous error correction, and boundary awareness, advising PMs to avoid over-relying on single metrics, neglecting user psychological costs, or

Stanford research finds AI has not caused mass unemployment yet, but impacts vary significantly. U.S. graduate unemployment reached 5.6% in 2026, with young workers and junior positions bearing the brunt, while senior employees remain stable. AI currently consolidates tasks and reduces new hiring rather than directly displacing workers, with enterprise adoption still in early piloting stages.

Yann LeCun's team releases VISReg, which solves the representation collapse problem in JEPA world models by decoupling regularization into independent "scale" and "shape" objectives. It outperforms 7 mainstream methods across 15 datasets, matching DINOv2's OOD performance using only ~1/10 of the training data, without any heuristic training tricks. Core implementation requires only ~15 lines of PyTorch code.

Peking University, Tsinghua University, Beihang University, Shanghai Jiao Tong University, and University of Science and Technology of China jointly launched TriWorldBench Challenge, introducing the first evaluation benchmark for robot tri-view world models. The benchmark contains 500 tri-view synchronized episodes covering 50 robotic manipulation tasks, evaluating models across 6 dimensions with 19 metrics to advance embodied world models from visual generation to world understanding.

Jensen Huang's second tweet announces the Open Secure AI Alliance with 37 founding members including Nvidia, Microsoft, and Meta, promoting open-source models for cybersecurity. Core argument: open-source models and harnesses are essential for defenders. Background: Hugging Face was invaded by AI agents, with OpenAI's closed-source models breaking security limits during attack tests while commercial models refused to analyze attack logs, ultimately completed by China's open-source model GLM 5.2.

SSI, founded by Ilya Sutskever, has formed a long-term strategic partnership with NVIDIA. NVIDIA will invest approximately $5 billion and provide the Vera Rubin computing platform, enabling SSI's computing power to increase 10-fold within 12 months. SSI is valued at $32 billion and has raised approximately $7 billion to date, having previously partnered with Google Cloud.

Show 1 related articles

Ilya Sutskever's startup SSI has established a strategic partnership with NVIDIA, which will invest $5 billion and provide substantial GPU computing power, enabling SSI to scale its computing capabilities by 10 times over the next 12 months. Founded in 2024, SSI focuses on developing safe superintelligence and was valued at approximately $30 billion last year, having previously raised $7 billion.

Lilian Weng, co-founder of Thinking Machines Lab, announces departure due to health exhaustion from months of overwork. Her farewell letter reveals frequent illness over the past 7 months, with health no longer sustainable for startup pace. She spent 20 months at the company, which recently launched its significant model Inkling.

A regular office worker has transformed an old Logitech MX Anywhere 3S mouse into an "AI accessory" using the Smart Actions feature in Logi Options+ software. By binding complex AI workflows\u2014select content, copy, invoke AI, paste\u2014to a single button, users can access ChatGPT, Typeless voice input, and other AI tools with one click without memorizing hotkeys. Advanced features include app-specific configurations where the same button triggers different actions in different software. Comp

Ant Group releases and open-sources LLaDA2.2, the world's first large-scale Agentic diffusion model, enabling diffusion language models to perform self-correction in long-horizon tasks for the first time. The model supports four atomic editing operations and scores 53.83 on average across 7 Agent benchmarks, delivering 1.64x the throughput of comparable autoregressive models with 128K native context support.

This article explains how AI agents perform web searches through broad searching and targeted extraction, using Doubao Search as an example to demonstrate the complete process of customizing a dedicated search Skill from scratch. Doubao Search provides 500 free queries monthly, with overages at 0.02 yuan per query and monthly plans starting from 0.005 yuan per query. Mastering these techniques enables more accurate and timely information retrieval for AI agents.

Frost & Sullivan report shows Baidu Miaoda ranks first in China's AI no-code platform market with 33.4% market share. Miaoda helps users complete the full chain from app creation to commercial monetization by providing app promotion, App generation, multi-terminal delivery and operations capabilities. It has served 35 million cumulative users, created 3.5 million commercially valuable apps, with 41.5% of users achieving actual revenue.

Anthropic's Claude independently completed AMD MI355X GPU adaptation over a weekend, bypassing Nvidia's 20-year CUDA moat with continuously improving performance. AMD launched ROCm.AI toolchain enabling AI to directly read chip documentation and invoke GPU resources. Anthropic announced plans to deploy 2GW of GPUs on AMD Helios systems, with initial 1GW targeted for first half of 2027.

Huawei and 6 universities jointly released Boogu-Image-0.1, trained with only 208M images at ~$400K, achieving open-source SOTA on multiple benchmarks. Supporting native 2K output, weights and code are Apache 2.0 licensed, available on ModelScope and ComfyUI.

Physical AI company Zhengqi Future completed three rounds of angel financing within 8 months, raising hundreds of millions of yuan, with investors including CDH Ventures, Linear Capital, and CATL-backed fund. Its core brand QUORRA focuses on D2D short-distance mobility robots, with its first product QUORRA X5 already in mass production and securing commercial orders from over ten countries. The company developed QUORRA DoorMind world model and FLAT motion execution platform in-house, building a

At YC Startup School 2026, Boris Cherny, creator of Claude Code, revealed that Opus 5 scored 30% on ARC-AGI-3 (previously only low tens of percent), and can run autonomously for weeks. Claude Code deleted 80% of its system prompts for Opus 5 and performed better. The model once rewrote Bun's entire codebase from Zig to Rust in 11 days using a single prompt, autonomously posting progress updates to Slack. This demonstrates that the core challenge with current AI has shifted from writing instructi

Aether AI and UC San Diego jointly propose CD-LAM (Causally Debiased Latent Action Model), which purifies scene information from latent action representations to address the issue of world models ignoring control instructions. Experiments show over 30% reduction in action errors, with post-training requiring only 3k-6k fine-tuning steps, approximately 10% of baseline training cost.

Testing reveals Opus 5's reasoning effort levels produce an inverted U-shaped performance curve, with medium outperforming max settings. Higher effort drives the model to "overthink"—in code tasks, it spontaneously refactors unrelated code, expanding minor bug fixes into complete pull requests. Anthropic recommends low/medium for straightforward tasks, reserving higher effort for complex, extended reasoning chains.

Researchers from Renmin University and Ant Group release SearchOS, an open-source framework providing operating system-level collaborative infrastructure for search agents. The framework includes context management, pipeline scheduling, tool middleware, and a skill library with around 280 skills. On WideSearch and GISA benchmarks, SearchOS achieves Set F1 scores 13.4 points higher than the best baseline, solving issues of state loss, redundant queries, and information gaps in long-horizon tasks.

Spectro Cloud has raised over $100 million in Series D funding, achieving a valuation exceeding $1 billion to become a unicorn. Goldman Sachs continues to lead the round, with participation from AMD, Ericsson and others. The company manages GPU resources across cloud, data center, bare metal and edge environments through its PaletteAI platform, improving enterprise GPU utilization from 30% to approximately 60%, with local inference reducing token costs by up to 70%.

AnySearch is a search infrastructure designed for AI Agents, built by a Chinese team of post-1995 and post-2000 members. In just two months, it has attracted 200,000 users and over 20 million API calls, with 6,000+ GitHub stars. Testing shows 76.4% accuracy, 12 percentage points higher than traditional search. It covers 20+ vertical domains including finance, legal, and academia. The intelligent intent routing automatically decomposes tasks and routes queries to appropriate data sources. Search

Moonshot AI announces the official open-sourcing of Kimi K3, the world's first 3-trillion parameter level open-source model (actual 2.8T parameters). K3 supports 1 million token ultra-long context with performance competitive against top closed-source models like Fable 5 and GPT-5.6 Sol, at half the inference cost. Open-sourced content includes model weights, technical report, and key Infra technologies like MoonEP.

Show 3 related articles

On July 17, Moonshot AI released Kimi K3 (2.8T parameters, 1M token context), fully open-sourcing model weights, technical reports and training infra on July 27. This is the world's first 3T-class open-source model. Jensen Huang then published an open letter opposing bans on open-source models, gaining support from 77 tech companies. The US government previously accused Kimi K3 of IP infringement.

MoonShot AI releases Kimi K3 with 2.8T total parameters, 104B activated parameters, and 1M token context. Key innovations include KDA for long-sequence state compression, AttnRes for cross-layer information selection, and Stable LatentMoE for efficient expert communication. Combined with partial rollout and microVM pause/resume capabilities, K3 enables 24-hour continuous agent execution with leading performance in kernel optimization tasks. Overall scaling efficiency improved 2.5x.

Kimi K3 officially open-sources with 2.8 trillion parameters, ranking first in open-source by a wide margin, receiving Day-0 support from vLLM, SGLang, and multiple platforms. The tutorial demonstrates integrating K3 into Codex client via Teamo tool, achieving API compatibility conversion and supporting model switching, completed in under ten minutes.

Tuchan Intelligence, together with Peking University and Peng Cheng Laboratory, released UniWorld-View, ranking #1 on Li Fei-Fei's WorldScore world model leaderboard, with Huawei Ascend adaptation and open-sourced code and weights. The model enables novel view synthesis from single images or videos, using occlusion-aware point cloud rendering to solve tearing and light leakage issues in large-baseline generation, supporting 360-degree orbital view generation.