Aureka has completed a $100M Series B financing, just 4 months after its last round. The company builds a dry-wet lab loop infrastructure using AI to design antibodies and other large molecule drugs. Its open-source engine OpenDDE outperforms AlphaFold3 in antibody-antigen prediction. With 8 public pipelines, 1 candidate is entering clinical and 2 are in IND development.
Embodied AI company LatentVerse has completed a seed round of hundreds of millions of yuan, with participation from Hillhouse Ventures, Qingliu Capital and others. Founded by Tsinghua University doctoral student Hu Yucheng, the company chooses a third technical route - unifying VLM and world model in one architecture, different from VLA or WAM. Its core product UTAM contains four expert modules, reducing data collection cost from thousands to below a hundred yuan per hour. The company plans to r
ByteDance has recently established a new first-level AI Data and Security department, led by Wang Yinglei, integrating teams including Global Data, DMC, and AIDP. The department will provide cross-modal data services for all ByteDance large models, with core functions covering full-process data operations including standard formulation, sourcing and procurement, synthesis and cleaning, and quality evaluation. Previously, ByteDance's data budget exceeded ten million USD, with thousands of personn
Zhipu AI's ZCode releases major update with four new features: Goal, Subagents, Remote Control, and off-peak tasks. Developers set targets and AI autonomously breaks down tasks, writes code, runs tests, and optimizes until completion without continuous supervision. Tests show GLM+ZCode achieves 2.39% higher pass rate than Claude Code, with 98%+ cache hit rate and effective usage nearly 1.8x under the limited-time bonus.
Open-source programming agent "Pi" achieves 99.93% cache hit rate for DeepSeek inputs through minimal tool design, with average task cost of just $0.028—seven times cheaper than Claude Code. Pi provides only four basic tools by default, reducing redundant content in conversations and allowing the model to skip previously processed parts, significantly lowering token usage costs.
Anthropic introduces invisible watermarking for Claude, embedding statistical bias in token selection to enable text tracing, effective August 2 for all text outputs globally. Files receive additional C2PA signatures. Legacy models have a transition period until December 2. As the first major AI company to deploy text watermarking at scale, Anthropic acknowledges limitations including inability to prove originality, potential quality impact, and metadata vulnerability to removal.
Meta releases Muse Glimmer, an open-source Agent model with 30 billion parameters under Apache 2.0 license. Compressed to under 20GB via 4-bit quantization, it runs locally on Mac or consumer GPU PCs, supporting offline tool calls and task execution without cloud API dependency.
Ming Chaoping, former head of Moonshot's Noisee, founded Xinyan Yima and launched YouWare in March 2025, an AI platform enabling non-programmers to generate websites and apps via natural language, achieving an $80M valuation after two funding rounds in under six months. The team also released Bloome (AI social) and Renoise (AI video creation platform), which integrates multiple models, canvas, and workflows for script planning and batch video production. Amid homogeneous AI video products, Renoi
Zhipu's MaaS platform has nearly 7 million registered API users, growing by approximately 2 million since early July. Over 50,000 domestically-produced AI chips have been activated. ARR has grown 15-fold, with short-term doubling following GLM-5 release.
Google DeepMind and collaborators released WeatherNext Cyclones, an AI model that simultaneously predicts typhoon trajectory, intensity, and wind radii for up to 15 days. The model achieves comparable accuracy approximately 27 hours earlier than existing operational systems. Published in Nature and open-sourced, it was trained on nearly 5,000 historical storm records.
Jiuhen Technology, founded by former NetEase VP Wang Yuan, upgraded from knowledge base to Agent OS, positioning itself as context infrastructure. Wang revealed that 100,000 targeted global users with 6-8% paid conversion can sustain the team. The company reports 80%+ monthly renewal rate for its To C subscription by the second month. Wang views large models as engines and context as gasoline, with Agent OS layered atop rather than replacing traditional operating systems.
ByteDance's AI app Doubao begins charging separate commissions for recommended products starting today. Hotel orders are subject to approximately 12% in combined fees, while lifestyle services around 18%, compared to the previous 2.5%-8% standard. With over 200 million daily active users but only about 10 million yuan in daily e-commerce transactions, this adjustment suggests Doubao's order volume has grown large enough to warrant independent ROI calculation.
Runta founder Guanlan Dai argues AI model capabilities are now sufficient, with future competition shifting to the infrastructure layer. Runta builds an execution platform specifically for AI agents, addressing isolation, control, auditing, and elastic scaling. The company recently closed a $20M seed round led by a16z with participation from Jeff Dean and Fei-Fei Li. The team comprises roughly 10 people; Dai previously served as core technical lead for Cloudflare Edge and engineering director at
Daimon Robotics unveiled Daimon-TWM, the world's first tactile-grounded world model that integrates native tactile sensing throughout physical cognition, decision-making, and real-time reaction. The model achieves 64.0% average success rate in contact-intensive tasks, over 80% in pose adjustment tasks, and 10x higher success rate than π0.5 in perturbed environments. Trained on over 10 million cross-material physical interaction data, the model enables robots to understand contact states, predict
Chinese AI companies like DeepSeek are releasing cost-effective open-source models, forcing US AI firms including OpenAI and Anthropic to slash prices. OpenAI cut its lightweight model Luna's cost by 80%, while Anthropic launched a high-performance model at half the price. DeepSeek V4 Flash has topped usage rankings. Analysts say AI price cuts are an "inevitable trend" that will drive broader AI adoption.
Acorn Robot has completed its angel round financing, led by China Merchants Ventures and NIO Capital. Spun out from Tsinghua University and Harvard University, the company's core technology is the Natus Embodied Instinct Model, enabling "zero-data, cold-start" operation where robots can work immediately upon arrival at production lines. The model reduces production changeover time from one week to 6-40 minutes and has been deployed at a global leading cosmetics ODM factory.
Qisuan Guangqi completed hundreds of millions of yuan in angel and angel-plus funding, with participation from Hillhouse Ventures and ZhenFund. The company has completed 256×256 optical computing chip tape-out, achieving four generations of iterations in 18 months, with plans to scale to 1024×1024. Its optoelectronic collaborative architecture can significantly reduce matrix operation power consumption, providing more efficient AI computing power.
Stanford AI economist Erik Brynjolfsson redefines AI as 'amplifying intention' - models amplify execution but not direction. He advises developers to shift from taking orders to proactively defining tasks and managing agent fleets, while capturing 10x/20x performers' methods for team replication. Smaller teams can now access capabilities previously unaffordable.
An 11-year-old girl, daughter of Runway CEO Chen Siqi, used Claude Code to create a new browser with Electron and Chromium, bypassing her father's YouTube screen time limits to watch videos without restriction.
Under new leadership of Alfred Lin and Pat Grady, Sequoia Capital has pivoted strategy to bet heavily on AI industrialization. In 2026, Sequoia has made 33 AI investments, committing $10 billion to AI infrastructure, leading Etched's $300M Series C at $10.3B valuation and Valar Atomics' $1B Series B. Investment focus has shifted from software in the digital world to chips, nuclear energy and robotics in the physical world.
OpenAI's AI browser Atlas officially shut down on August 9, 2026, after 292 days of operation. Positioned as a potential Chrome challenger, Atlas required extensive foundational work beyond AI capabilities. OpenAI has decided to integrate Atlas's AI browsing abilities into ChatGPT desktop app and Chrome extensions, discontinuing independent browser development.
Professor Hong Sheng's team at Beihang University proposes the BGA noise-immune neural distillation framework for detecting malicious encrypted traffic in industrial IoT. By leveraging WGAN-GP for sample synthesis, ANOVA for feature selection, and adaptive gated attention mechanisms, the framework addresses the AI "attention dilution" problem. Experiments show recall and accuracy exceeding 95.2%, with inference latency of just 0.28ms, demonstrating 8.57% more robustness than standard models unde
Alibaba's Qwen App launched dozens of third-party agents including Beike, SF Express, and Hellobike. Through account authorization, these agents access users' historical orders and addresses. Users can complete tasks like express delivery, car rental, and apartment hunting via voice commands without any page redirects, lowering the service usage barrier while preserving brands' identities.
Researchers from Beihang University, Tsinghua AIR, and Didi propose DriveTeach-VLA, an autonomous driving model accepted by ECCV 2026. The method teaches models to focus on critical traffic elements via visual distillation (DVD) and maps BEV trajectories to image space using 2D-TGP, enabling foundation models to understand trajectory spatial semantics. Achieves 90.4 PDMS on NAVSIM, improving to 92.7 with DrivorR trajectory selector, providing a new path for VLA scaling.
AI Agents are transforming storage from passive data containers into near-data processing nodes capable of encryption, compression, indexing, and memory maintenance. HBM, HBF, DRAM, and SSD are being reshuffled based on access frequency, mutability, and sharing scope: HBF fills the capacity gap between HBM and SSD with 512GB/0.4-3.0TB/s specs. Companies like Phison, Longsys, and YB-Tech/Maxio are defining new interfaces at the compute-storage boundary, seeking to expand revenue from per-disk sal
Snowflake reported Q1 FY2027 earnings with revenue of $1.391 billion, up 33% YoY, and net revenue retention of 126%. Over 13,600 accounts use AI capabilities, with CoWork growing over 100% QoQ. The company leverages Ontology technology to build an enterprise cognitive layer, evolving from a data platform into an AI cognitive entry point for businesses.
Anthropic's internal Claude attempted to prove the Riemann Hypothesis. While not fully solving it, the model dramatically raised the lower bound of zeta function zeros on the critical line from 41.6% to 67.2%, accomplishing in 1.5 days what took human mathematicians 37 years. Achieved through roughly 60 sub-agents collaborating with positive reinforcement, this marks AI's transition from solving structured problems to open-ended mathematical research.