Covering 1054 days, 共 26880+ articles built agent-first, open to all tech stacks.
Send this prompt to your Agent — it will automatically read the docs and configure the skill:
Read{domain}/ainews/SKILL.md ...
According to the UK's Financial Times, DeepSeek has started preliminary discussions for a second round of funding this week, with a pre-money valuation of about $71 billion (approximately 481.4 billion yuan). The company plans to launch the official DeepSeek-V4 model this month, and the funding will be used to expand computing power infrastructure.
Jenny Wen, Claude's core design lead, has resigned to join Cursor as Head of Design. Cursor's parent company Anysphere was recently acquired by SpaceXAI, meaning Jenny is effectively joining Musk's team. At Anthropic, she led Claude.ai's visual redesign and the Cowork general agent product, and once declared that traditional design processes are dead in the AI era.
Richard Sutton, the nearly 70-year-old Turing Award winner and father of reinforcement learning, co-founded Oak Lab with Khurram Javed, aiming to develop OaK, an autonomously continuous learning AGI architecture, with the goal of building a trillion-parameter agent that learns in real-time with only 20 watts of power consumption.
Kuaishou releases flagship Agentic Coding model KAT-Coder-Pro V2.5, which can access Claude Code. Its long-range engineering capability is close to Opus 4.8, can independently complete complete programming projects, and is now fully online on StreamLake.com.
The AI film 'Zhuangzi's World: Butterfly Dream', produced by scholar Yi Zhongtian and directed by Liu Shuwen, premiered exclusively on Douyin and surpassed one million views in three days. Adapted from 'Zhuangzi: Discussion on Making All Things Equal', the film uses AI to reinterpret the classical allegory in a sci-fi style. Production took six months and consumed 1.2 million tokens, featuring the first live symphony orchestra performance for an AI film. Yi Zhongtian revealed that the team is de
AI video unicorn Aishi Technology has completed a 2.98 billion yuan Series C funding round led by Alibaba. The funds will be used for video generation foundation model and real-time world model R&D, as well as global expansion. Its founder Wang Changhu is a former head of visual technology at ByteDance.
Vertex Lab's V2Fun becomes the first native 3D large model AI application for HarmonyOS. It supports generating 3D models that can be exported and 3D printed from a single photo within tens of seconds, lowering the threshold for 3D creation, topping Huawei AppGallery's editor's choice and entering the download chart's Top 30.
Elon Musk admitted that SpaceXAI's Grok Build coding tool secretly transmitted user code, and has ordered deletion of all historical user data. A security researcher discovered the tool uploaded 5.1GB of test repository data, with the data collection toggle ineffective, and it was revealed to have once uploaded an entire computer home directory.
Xense Robotics completes a 100-million-yuan funding round, jointly invested by a top embodied intelligence industry player, Jide Electric, and Tianji Capital. The funds will accelerate the industrialization of high-precision tri-color optical vision-tactile sensors, deploy tactile data acquisition equipment and datasets, and refine its physical intelligence model natively supporting tactile modalities.
Anthropic research reveals that Claude's value orientations vary by model version and conversation language. After analyzing 300,000 dialogues, the team proposed four value axes: accommodation vs. caution, warmth vs. rigor, depth vs. brevity, and candor vs. execution. Sonnet 4.6 is warmer and more concise, while Opus 4.7 is more rigorous and cautious. Claude is more accommodating in Arabic conversations and more cautious in English.
LibTV Agent has rolled out a major update, integrating 100+ professional AI video Skills to become the world's largest Skill Hub, covering styles from directors like Wes Anderson, Wong Kar-wai, and Pixar. Users can generate professional-grade videos with one-click natural language input and create personal original Skills to capture their methodologies. The platform has invested 10 million yuan to incentivize creators, combined with Agent-driven post-production adjustments, lowering the AI video
Holi-Spatial developed by a team from Shanghai Jiao Tong University has been accepted as ICML 2026 Oral. This technology requires no manual annotation nor LiDAR, it can automatically generate 3D spatial annotation data using only ordinary videos, achieving 81.06% 3D detection accuracy on ScanNet++, outperforming existing methods by a large margin.
Sam Altman launched a GPT-5.6 Sol work collection event on 𝕏, netizens submitted various application works. The model's token efficiency in Agentic Coding tasks has increased by approximately 54%, OpenAI has removed the 5-hour usage limit for Codex packages and reset user quotas.
Fudan University releases the 2026 Blue Book on AI Development in Humanities and Social Sciences, proposing a shift from one-way empowerment to bidirectional integration between AI and humanities/social sciences. It emphasizes that the stronger AI becomes, the more important deep thinking becomes; research evidence chains cannot be outsourced; AI governance must cover the full lifecycle; and humanities/social sciences should determine why and in what direction AI is used.
Self-supervised learning method VISReg, endorsed by Turing Award winner Yann LeCun, solves the representation collapse problem of JEPA world model by decoupling regularization into scale and shape components. It outperforms 7 mainstream methods on 15 datasets, matches DINOv2 with 1/10 of the training data.
Anthropic announced that Claude Artifacts now support public sharing and multiplayer collaborative editing. Artifacts created by Claude Code can now generate public links, and team members can jointly edit and maintain the same deliverable. Meanwhile, Claude Tag in Slack can directly create Artifacts, converting channel discussions into deliverable pages, linking the complete collaboration chain from task delegation to result delivery.
Tencent's WeChat AI team has spent four years developing WeLM, a large language model using an 80B-A3B sparse MoE architecture trained on under 14T tokens, with KV-Mirror and MTP layers optimizing inference speed. The team pioneered Hidden Decoding latent space computation technology, scaling WeLM-HD4 to 617B parameters and outperforming autoregressive baselines. The model powers WeChat's "Xiaowei" assistant via a hybrid "WeLM+DeepSeek" architecture and is open-sourced on GitHub and HuggingFace.
Microsoft CEO Satya Nadella proposes the "Reverse Information Paradox" in AI: enterprises pay twice for AI usage—one payment to model vendors, and the other is giving away their proprietary knowledge. He puts forward solutions advocating enterprises should control their own knowledge assets.
Inspur launched the industry's first CPU native liquid-cooled whole cabinet server and Yunniao SD200 super-node AI server at OCTS26, which can support more than 40,000 agents to collaborate simultaneously, and the single Token generation latency of Yunniao SD200 on trillion-parameter models is as low as 4.77ms.
A tester conducted a practical comparison test of three Chinese desktop Agents—ByteDance Doubao Pro, Alibaba QoderWork, and Tencent Cloud WorkBuddy—across 8 core capability dimensions using 8 real office tasks. Each product has clear advantages and disadvantages, and users are suggested to try the free quota before paying.
Looki (Guangzhi Space-Time) completes hundreds of millions of yuan in Series A1 financing led by Genesis Capital, with existing shareholders following on. The funding will be used for multimodal model R&D, product matrix construction and team expansion to accelerate the development of personal AI recommendation engine for the physical world.
Reve has released version 2.1 of its image model, ranking second on the Arena text-to-image leaderboard, only behind OpenAI's GPT Image 2. Its training compute is only one-tenth of competitors, supporting more flexible image editing and natively outputting 4K 16-megapixel images.
MOVA LINCO has completed tens of millions of yuan in angel funding, focusing on home AI infrastructure. Its product lineup includes the AI voice router X1 Pro, AI NAS UP6, and AI Box, building an integrated architecture of connectivity, storage, and computing. The X1 Pro supports natural voice interaction, with the overseas version launching in late 2026.
General humanoid robot company LimX Dynamics completes $200 million Pre-IPO financing, with a post-investment valuation of 15 billion yuan. 70% of the capital comes from overseas, and it is currently preparing for a Hong Kong stock IPO. The funds will be used for technology R&D, robot delivery and global layout.
In a conversation with LangChain CEO, Jensen Huang stated that LLMs have become infrastructure and future companies will be built on Harnesses, with core competitiveness being proprietary AI agents. Nvidia engineers prefer building Agents to writing Python, and open-weight models have approached frontier performance at just one-tenth the cost.
Japanese physicist Yuji Tachikawa shared: a research project stalled for six months found a breakthrough overnight with Claude Fable 5. The model found a calculation error and proposed new approaches. Anthropic extended Fable 5's open access until July 19.
Developer Eric Lu released Ghost Font, which leverages pixel motion direction differences for instant human comprehension while AI fails completely, amassing 17 million video views. However, just one day later, prompt engineer Riley Goodside broke through GPT-5.6's defenses with a single directional hint to read the hidden message, proving the gap between AI and human perception is just one prompt away.
Three Anthropic executives discussed the future form of Agents, noting that Claude Platform has evolved from a pure API into a managed agent platform integrating memory, identity, permissions, and sandbox simulation. Improved model capabilities have thinned the scaffolding, making multi-agent collaboration strategies mainstream. Enterprise ROI should start with individual efficiency, then scale to teams and cross-team workflows. In the future, agents will be deeply embedded in organizations like
Terence Tao used an AI agent to migrate his personal website built in 1997 within one day, transferring data including 560 papers into a new tao-web repository. The AI also ported his Java applets from over 20 years ago to JavaScript and identified two unknown bugs in his old code.
Nine top talents joined Anthropic in the first half of 2026: Nobel laureate John Jumper, AI educator Andrej Karpathy, YC partner Tom Blomfield, xAI's last co-founder Ross Nordeen, Stanford economist Chad Jones, and Berkeley CS chair Jelani Nelson, spanning AI research, enterprise, philosophy, and economics—betting that AI will become the core infrastructure of the next decade.
Peking University's Yang Yuchao team, in collaboration with a Chinese Academy of Sciences team, has developed the world's first memristor neurodynamic chip based on controllable in-memory computing. The single-step delay is compressed to 2.12 milliseconds, with significantly improved speed and reduced power consumption compared to existing chips, providing a new hardware foundation for fields such as brain-computer interfaces.
xAI releases Grok 4.5 with 1.5 trillion parameter MoE architecture, 500K context window, and output pricing at 1/4 to 1/5 of competitors. Testing across 7 coding tasks shows performance close to Opus on basic coding tasks with fast generation and high cost-efficiency, though gaps remain in aesthetic judgment and creative expression scenarios.
A team from Fudan University releases LEGALWORLD, an LLM agent-driven legal simulation environment. It contains over 70,000 simulated cases covering more than 500 civil causes of action, supports full-life-cycle case推演, and can be used for evaluation and training of legal agents.
Peking University and DP Technology propose Parallel Rollout Approximation (PRA), which addresses bottlenecks in pixel-space autoregressive image generation via low-dimensional intermediate states and parallel rollout approximation, eliminating the need for visual tokenizers. The 135M model achieves FID 2.58, surpassing the 1.9B-parameter FARMER baseline, while the 511M model reaches a new state-of-the-art FID of 1.94 and demonstrates strong visual understanding capabilities.
md2wechat releases version 3.0.0, focusing on 10 image layout scenarios including multi-image grid, sequential storytelling, mobile screenshots, image-text combination, key annotations, before/after comparison, step-by-step tutorials, formal captions, and two SVG interactive effects, enabling ordinary creators to complete multi-image and SVG layouts for WeChat official accounts via Agent. The tool has gained 3.2k stars on GitHub.
On July 13, StepFun launched the agent-native AI terminal brand STEPX, the world's first agent-native operating system Step AOS, and the new personal agent Amoo in Shanghai. The world's first large-model-native agent phone STEPX Neo debuted simultaneously and has passed the national standard L3 test.
Since April this year, at least 11 core executives and key project leaders have left OpenAI, including the AGI Deployment CEO, COO, and CFO. OpenAI is undergoing a strategic transformation, shutting down side projects such as Sora and smart hardware while concentrating resources on the enterprise market and commercialization.
Tibo, head of the OpenAI Codex team, publicly guided users on integrating the GPT-5.6 Sol model into Claude Code via CLIProxyAPI (a GitHub project with 41,000 stars), providing the alias "claudex." The situation arose because Codex's multi-agent capabilities lag behind Claude Code, prompting the community to combine them through a reverse proxy tool. Tibo also corrected a Chinese community user on misconfigured token-saving settings and rolled back the context window from 372K to 272K. In an ear
Kuaishou's Kro team launched KroWork, a desktop AI agent that lets users generate runnable applications via natural language and pin them as desktop shortcuts for reuse without spending credits. The domestic version integrates DeepSeek v4 pro and Qwen 3.7 Max, while the international version supports OpenAI and Anthropic models. Testing shows a high first-pass success rate for basic tasks, but occasional PPT generation interruptions, UI rendering lag, and slow browser automation.
Open-source project Mindwalk releases v0.1.0, which reads local session logs from Claude Code and Codex and replays AI coding assistants' work paths on a 3D code map. It supports Tree and Terrain views, marks file access status with colors, and provides metrics such as error rate and churned files for developers to review Agent task execution.
Zhang Jintao, a post-2000-born PhD student of Tsinghua Professor Zhu Jun and the end-to-end lead of Vidu S1, drives inference optimization efforts including SageAttention, TurboDiffusion, and TurboServe. These works underpin Vidu S1's real-time streaming video generation at 540P and 40FPS, supporting voice-controlled full-body motions, voice cloning, and consumer-grade GPUs, with its API now open to the public.
AI Agent startup Lyzr deployed its self-developed Agent SivaClaw for outreach during its Series B funding. The round is expected to raise $100 million at a $500 million valuation, attracting $400 million in investment interest. Clients include Accenture, Deloitte, and KPMG.
CB Insights releases 2026 AI 100 list, selecting 100 AI startups from over 40,000 global companies. The selected companies have raised a total of $10.9 billion. This year's winners focus on three major areas: AI Agent governance and security, Physical AI infrastructure, and building data barriers in professional industries.
Yuxun Network, a Tsinghua-affiliated startup, completed a tens-of-millions-yuan funding round led by Zhipu, focusing on its ZCube flat network architecture. Verified in Zhipu's thousand-GPU cluster: 15% higher Token output, 40.6% lower latency, and 33.3% reduction in network hardware costs, saving 210-640 million yuan on ten-thousand-GPU clusters. The company has secured over 100 million yuan in orders.
DeepSeek adopts two rounds of strict code interviews for programmer recruitment. All questions come from online actual combat, cannot be answered through traditional question-memorizing routines, and adheres to "code is justice". The interview lasts about 70 minutes, and the third interview assesses cognition and judgment.
MiHoYo's AI companion digital human product "BSide: Olivia Lin" has opened early access today. It combines AI companion, music playback and dynamic wallpaper, supports letter interaction and user-uploaded MIDI files, and the test will last until the end of 2026.
