Global AI Daily News

SK Hynix listed on Nasdaq on July 10, 2026, raising $26.5 billion in its IPO, the largest ever, with a market cap exceeding $1 trillion. Through strategic acquisitions of energy and telecom monopolies via political connections, the company bet on HBM technology, transforming from an also-ran to the most critical AI memory supplier. Currently holding 58% HBM market share, SK Hynix is deeply integrated with Nvidia to co-develop next-generation memory solutions.

OpenAI launched a unified search for ChatGPT on July 14, supporting searches across chats, images, documents, and projects, synchronized across Web, iOS, and Android platforms with all plans globally available. This addresses the long-standing pain point of finding old conversations over the past four years, enabling ChatGPT to search its own content for the first time and evolving from a chatbot into a personal knowledge base.

SenseTime unveiled SenseNova-U1 Pro at WAIC 2026, the first industry solution integrating "seeing, generating, and acting" into a unified multimodal base. Powered by its proprietary NEO-unify architecture supporting native 8K output, it leads in text accuracy and Chinese aesthetic detail, surpassing GPT-Image 2, positioned as a "deliverable" creation model enabling the upgrade from "can generate" to "can deliver".

Shanghai AI Lab proposes Self-Harness, enabling the Agent's outer harness to automatically improve based on its own failure trajectories without changing the underlying model. Experiments show Qwen3.5-35B-A3B achieves 104% improvement, while MiniMax M2.5 and GLM-5 gain 28% and 24% respectively. The method employs a three-step closed-loop process of weakness mining, modification proposal, and regression testing to prevent arbitrary changes.

TeleAI unveiled AI Flow at WAIC 2026, enabling direct satellite connectivity for lightweight drones to transmit HD video in real-time. By encoding video into Token sequences for transmission and reconstructing via generative models, this technology addresses bandwidth constraints. Voice calls now operate at 0.266Kbps, video at 100Kbps, with storage reduced to 1/40 of conventional solutions, already deployed across 31 provinces for flood control.

Tec-Do Technology unveiled Navos 2.0 at WAIC, upgrading to an Agentic Workflow architecture featuring five core agents: AI Product Selection, Creative, Influencer, Manga, and Desktop Assistant. Their self-developed Tec-Chi model scored 85.82 on SuperCLUE-Mkt, ranking first globally. The company operates 2,000-3,000 agents internally and has partnered with OpenAI on A2A Agentic Commerce. A post-95s entrepreneur achieved $10,000+ GMV using Navos, validating the one-person company model.

Berkeley postdoc Haven Feng releases [schema] framework, achieving 98.98% RHAE on ARC-AGI-3 public set. The framework enables AI to work like a physicist by writing its understanding of the world into executable programs, then validating hypotheses and correcting errors through experiments, without modifying the model itself.

Tencent News head He Yijin delivered a speech at WAIC 2026, stating that Tencent News is advancing AI applications in the news industry through its 'News + AI' strategy, improving content creation and review efficiency on the production side, while redefining news consumption through 'News Girl' Q&A and 'AI Radio' voice interactions on the consumer side, and building an AI middle platform to consolidate reusable capability assets. Tencent Technology simultaneously released the 'Q2 2026 Global AI

At WAIC 2026, Qingwei Intelligence showcased a complete ecosystem for domestic AI chips, using reconfigurable architecture to break through manufacturing constraints and building system capabilities through 4K super nodes and RAISA software stack, achieving deployments exceeding 5,000 PFLOPS and adapting over 200 models, with 4K super node interconnect costs reduced by 90%, marking a shift in domestic AI computing competition from single-chip parameters to scalable infrastructure capabilities.

At WAIC 2026, Yuanli Lingji partnered with StepFun to complete a 3.5-meter Great Wall model using 8+ components with 6 robots operating continuously for 15 hours. This demonstration, based on the combination of base model DM0.5 and world model DW0.5, achieves sub-millimeter precision operations and long-horizon task execution, validating the embodied AI approach of combining foundation models with post-training paradigms.

Anthropic announced that starting July 20, Fable 5 will permanently remain in Max and Team Premium plans with a 50% rate limit, while Pro and Team Standard users receive a one-time $100 credit. Previously, Fable 5 faced a 19-day compliance ban and was scheduled for removal. Kimi K3's release and GPT-5.6 Sol's pricing at half of Fable 5's cost are viewed as key factors driving this policy reversal.

Former Google DeepMind scientist Alex Turner resigned and exposed Google's secret military AI deal with the Pentagon. He drafted a 25-page safety framework and collected signatures from approximately 250 employees, but senior leaders including Jeff Dean and Demis Hassabis ignored his pleas. Google ultimately signed an agreement allowing Gemini deployment in military data centers, completely eliminating AI chain-of-thought monitoring. Turner chose unemployment over lucrative job offers.

OpenAI CEO Altman publicly acknowledged poor performance over the past 12 months while predicting the most exceptional 12 months ahead. OpenAI and Anthropic are fiercely competing for users: Codex and ChatGPT Work gained 3 million active users in just four days, growing from 6 million to 9 million. OpenAI removed the 5-hour usage limit and reset quotas, while Anthropic extended Claude Fable 5 access and increased Claude Code weekly quotas by 50%. Both companies launched a new metric called "Usef

Shisec Technologies launches Vectron, a domestic Token optimization platform, compatible with 10+ domestic chips including Ascend and Kunlun, achieving 100 billion Token daily throughput. By leveraging technologies like KV Cache compression and long-context optimization, Vectron maximizes computational efficiency, addressing the underutilization issue in intelligent computing centers and unlocking trillion-level value from computing power.

At WAIC 2026, Turing Award winners John Hopcroft, Yin Qi, Zhou Bowen, Richard Sutton and other experts shared insights on AI's future. Key topics include: talent cultivation as core AI competitiveness; the next AI boom from agent models combined with terminals; scientific research as the next coding; experience-driven approach surpassing static data labeling; physical intelligence solving LLM hallucinations; and scientific data as the key to AI transformation.

BMPI and partners open-source StaffDeck, a digital employee construction platform that positions AI as formal corporate employees rather than chat tools. The platform addresses AI workflow chaos, answer credibility, and inability to self-evolve through process-oriented skills, OKF structured knowledge base, and feedback iteration loops. The Preview version offers 5 ready-made digital employees with support for private deployment and scheduled tasks.

Starchain Intelligence releases Lumo-2, its second-generation embodied foundation model, showcasing 20+ difficult household tasks via real robot demonstrations including dual-arm gift-wrapping, egg-flipping, and coffee-making. The model employs a 'predict-world-then-generate-action' approach, achieving 2.71x faster inference. Combined with Agent Philia and the T1 robot (starting at 89,900 yuan), this forms a full-stack architecture of AI model-embodied OS-tendon-driven body.

QiXun Intelligence showcased its Spirit v1.6 embodied AI system at WAIC, featuring an integrated VLA and world model architecture that enables long-horizon task execution. Moz1 autonomously decomposes tasks and dynamically replans when encountering environmental changes, such as new objects during 'living room tidying', achieving 99% success rate in CATL production lines. Debuting Moz2 targets commercial service scenarios, marking progress in the 'industry first, then commercial, then home' road

RoboScience Machine Science unveiled Visics at WAIC 2026, the world's first cloud-deployable embodied AI large model, achieving the industry's first "30-second hand-switching, plug-and-play" capability. Based on VLOA architecture and Object Trajectory Token design, it solves cross-body generalization challenges, reducing data acquisition costs to 1/20-1/200 of traditional solutions, with multi-SKU grasping success rate exceeding 99%. The partnership with Tencent Cloud launches EaaS services, pus

Anthropic releases Claude for Teachers on July 14, offering free access to US K12 teachers for lesson preparation, curriculum design, and learning analysis, valid until June 2027. OpenAI launched ChatGPT for Teachers in November 2025, with both AI giants competing for the education market. Domestic companies like TAL are rapidly deploying teacher AI tools.

Alibaba DAMO Academy open-sourced RynnWorld-Teleop, using a generative world model to replace real robots for training data collection. The system achieves 40FPS real-time inference speed, validates zero-real-machine training on dual-arm dexterous hands with 82.86% Block Pushing success rate, and combining synthetic with real data consistently improves task success rates.

Google and collaborators introduce VGGRPO (ECCV 2026), a geometry-aware video post-training framework that leverages implicit geometry modeling and two reward mechanisms in latent space, enabling camera motion smoothness and cross-view geometric consistency without architecture modifications or pixel-level rewards, effectively enhancing world consistency for both static and dynamic video generation.

FreeStyle introduces a community LoRA mining pipeline, constructing approximately 619K style reference sequences and 480K dual-reference sequences, with a two-stage training strategy using attention-level enrichment constraints and frequency-aware RoPE modulation to address content leakage in style transfer, along with a multi-dimensional evaluation benchmark.

The RoboDojo team releases RoboDojo, a unified embodied intelligence benchmark covering 42 simulation tasks and 18 real-world robot tasks, evaluating 30 representative robot manipulation policies. Results show the best current model achieves 8.80% success rate in simulation and 12.8% in the real world, compared to 76.03% and 100% for human experts respectively. This reveals a significant performance gap between current robot models and human experts, providing a common reference for subsequent m

This article presents a comparative test of three domestic office AI agents—QClaw, WorkBuddy, and Baidu Companion—across three scenarios: PPT generation, data visualization, and self-media content creation. Baidu Companion ranked first overall with the lowest subscription price, making it the most suitable choice for individual users' frequent office tasks.