Global AI Daily News

FDE (Frontier Deployment Engineer) has become the hottest new AI role, stationed at client sites to help deploy large language models. Silicon Valley salaries start at $170K annually, with top domestic tech firms offering up to $1M yearly, and individual contractors earning over $1M monthly. However, clients struggle to articulate pain points, system migration risks are high, and cost reduction space is limited, suggesting individual FDEs may be a transitional phase as enterprises develop intern

A developer implemented Kimi K3 (2.78T parameters) CPU inference using just 200KB of C code, reducing peak memory to 8.24GB through sparse expert architecture activating only 3.7% of parameters combined with weight streaming. Storage requires 1.56TB NVMe SSD, but speed reaches only 0.03 Token/s. DeepSeek V4 Flash, after quantization, runs on devices with 110-168GB RAM, marking frontier models entering personal desktops.

Sand.ai open-sources MAGI-2-preview, the world's first 100B+ parameter MoE video generation model, with 114B total parameters and ~6B activated per forward pass. Generating 10-second 1080P video costs only 0.5 RMB, roughly one-tenth of industry standard. Ranked sixth on AA video leaderboard, approaching top-tier performance. The model employs Multi-Head LatentMoE architecture, unifying video, audio, and text in a single stream, with proprietary infrastructure solving long-sequence and cross-devi

InSpatio open-sources QuerySplat (Topos-Lite), a feedforward 3D Gaussian Splatting technology that generates freely navigable 3D scenes from seconds of processing using just a few uncalibrated photos. Topos-Lite achieves state-of-the-art results across multiple view settings on the DL3DV-Evaluation benchmark in PSNR, SSIM, and LPIPS metrics.

Acti, developed by a Singapore team, is an AI-powered keyboard that allows users to query information, find files, and create meetings directly within chat by long-pressing the spacebar, eliminating the need to switch apps. It supports connections to 6 application types including Gmail, Slack, and Notion, with access to over 150 APIs and coverage of 95 languages. The product has secured a $5.3 million seed round led by BITKRAFT Ventures, with users having already created over 1,000 custom skills

Wenxin Assistant topped the SuperCLUE XClaw evaluation with 97.62 points, scoring 100 in memory, 98.86 in data processing, 99.44 in content creation, and 96.44 in research analysis. It achieved 94.6% on the global PinchBench v2 benchmark, ranking first among 59 models. The product remains completely free with no usage limits.

Xspark AI, together with top universities including Tsinghua, Peking University, HKU MMLab, and HKUST Guangzhou, released the TEI T0-T5 trustworthiness grading system (a 37-page survey paper), adding a new dimension to Physical AI evaluation. This grading system, benchmarking against SAE autonomous driving levels, proposes the industry's first quantifiable trustworthiness evaluation standard, expected to fill the long-standing gap in standardized assessment benchmarks and shift embodied AI devel

Yu Kaicheng from Westlake University founded Awomo, proposing a "conceptual world model" approach that establishes physical concept representations in abstract mathematical latent space instead of generating pixels, solving generalization challenges in physical intelligence. After rejecting his most successful work BEVFusion in late 2024, he pivoted to implicit methods. Founded in January 2026, the company has completed over 100 million yuan in seed and angel rounds with investors including Inno

Domestic AI model vendors are selling services at 90% gross margin with 10x cost pricing, while monthly packages costing 200-700 yuan are depleted extremely quickly, leaving users with little value. In contrast, overseas products at $30-100 are much more durable. High pricing squeezes application-layer viability and hinders AI adoption. The article calls on vendors to lower margins and make AI accessible to more users.

Musk reshared the claim that source code is becoming the next assembly language, advocating for AI to directly generate binaries and skip source code entirely, sparking industry debate. Clean Code author Bob Martin suggested keeping an intermediate layer, while UML creator Grady Booch argued that any precise description of intent is essentially a programming language. In April 2026, Google generated nearly 80% of new code with AI, yet still produces source code rather than direct binaries.

Glow raises $180M Series A at $1.2B valuation, becoming a unicorn. Founded by former Meta and Snowflake executives, the company is building an AI-powered endpoint security platform to help enterprises monitor software and AI agents on employee devices, assess risks in real-time, and block malicious programs to defend against AI-generated cyber attacks.

University of Southern California researchers released ActiveVision, a benchmark testing AI models' active visual reasoning capabilities. Results show GPT-5.5 scored only 10.6% versus human average of 96.1%—a 9x gap, while Fable 5 achieved only 3.5%. Even with tool-augmented agents, the highest accuracy reached only 50.6%. The core issue: models can "see" images but cannot continuously observe and verify—revealing fundamental separation between perception and reasoning.

Cloudflare launches programmable wallet Cloudflare Wallets for AI agents, providing readable usernames and virtual wallets. The virtual wallet sets a loss limit, enabling AI to autonomously complete payments, identity verification and commercial transactions within a controlled budget, eliminating the need for full human intervention and solving the challenge of AI's inability to self-pay.

Zhang Fan at the 2026 Singularity Intelligence Product Conference stated that AI era product managers' core competencies have shifted from "drawing flows, refining interactions" to "building the chessboard, setting directions, establishing boundaries." AI's native first principle is probabilistic rather than deterministic, with the core challenge lying in defining and constructing validators. Products must transition from "solid" to "liquid," with disposable interfaces becoming a new trend.

Two former OpenAI and Google core researchers Jerry Tworek and Rohan Anil founded Core Automation, arguing that the current AI approach has fundamental flaws: reinforcement learning cannot bring AGI, and Transformer models can only learn in laboratories, unable to continuously adapt to the real world after deployment. The two are developing next-generation architecture that enables AI to accumulate experience and improve continuously during use, eliminating dependence on repeated human retrainin

Anthropic CEO Dario Amodei worries that the company's million-dollar salaries attract talent chasing money rather than the company's "AI safety" mission, making it hard to distinguish genuine commitment from rehearsed answers. This reveals a "loyalty dilemma" among leading AI companies: high pay gets talent in the door, but cannot guarantee long-term commitment. Top researchers value compute resources, influence, and the freedom to choose technical directions.

Xu Lvyang, product operations lead at MiniMax Hub, went viral after appearing on Tencent Video's reality show "Half-Cooked恋人5". She embodies the dual pressures faced by professional women: capable at work yet expected to be soft in relationships. Zhaopin.com's 2026 survey shows 60.9% of female job seekers face fertility questions, with 32.1% of women aged 31-35 experiencing age-related career anxiety. Her portrayal of "high-functioning but running on empty" resonated widely.

At a San Francisco AI startup pitch event with 400 attendees, 8 teams showcased vertical AI applications: Cure Money won first place for AI financial assistants serving credit unions, while Atelier Studio reduced anime production costs tenfold with AI. Investors focused on competitive moats and paying users, revealing AI startups are shifting from competing on technology to competing on industry understanding.

Tencent Hunyuan releases Hy ASR 3.0 preview, a new speech recognition model based on large model Hy3, achieving comprehensive improvements in general recognition, dialect coverage, and scene robustness. On open-source benchmarks, Mandarin WER reaches 3.34%, English WER 2.62%, and Cantonese WER 3.12%, outperforming competitors. Supporting 10 major dialect regions, it can be applied to intelligent customer service, voice search and other scenarios.

MiniMax releases H3 video model at one-third the price of Seedance 2.0, with approximately 20% performance gap. Its core strength lies in video post-processing capabilities—understanding original video editing logic for precise modifications rather than simple face-swapping. H3 positions itself as a post-production tool rather than a generation tool, using an open-source strategy to cultivate user habits and carve out a differentiated market position.

Chinese flagship AI models are accelerating their embrace of native multimodal capabilities. Kimi's K3 uses vision as the Agent's core perceptual system, directly understanding web screenshots. Meanwhile, DeepSeek V4-Flash improves coding performance through post-training alone, without expanding parameters or adding vision. These two approaches reflect the deeper divergence among AI companies regarding the timing and tradeoffs of multimodal investment.

At UC Berkeley, Google DeepMind CSO Jasjeet Sekhon revealed Google's 2024 AI capex of $195-205 billion, with Silicon Valley betting on RSI (Recursive Self-Improvement). RSI means AI can independently redesign its own architecture, driving marginal costs toward zero. Anthropic's 9 Claude agents recovered 97% performance gap in 800 hours, while OpenAI's GPT-5.6 Sol cut inference costs by 20%. Sekhon warned of an "AI air pocket" risk but called it the largest scientific bet in human civilization hi

Sheng Ying, former inference team leader at xAI and initiator of open-source inference framework SGLang, has founded RadixArk, securing $100 million in seed funding to democratize cutting-edge AI infrastructure for the entire industry.

Anthropic signed a 6-year compute agreement worth $10 billion with AI cloud startup Volta Infra, which was founded just 7 months ago. Volta operates on a model of securing contracts before building facilities, using Bitdeer for Norwegian data center sites and power, while Nvidia supplies chips. With a valuation of only $2.4 billion, Volta secured a deal worth over 4 times its valuation. Anthropic has now signed cumulative compute agreements exceeding $100 billion this year.

DeepSeek releases Agent Harness beta recruitment, requiring developers to submit open-source Agent projects. 769 applicants contribute 712 repositories with 1.2M Stars across 18 tracks. This evolves into an open-source ecosystem inventory, with developers focusing on long-task execution, context and memory management, and security evaluation—outlining the evolutionary roadmap for AI Coding Agent engineering.

OpenAI is shutting down its Atlas browser just nine months after launch. AI browsers once hyped as the next internet gateway are failing due to high user migration costs. Gemini and Claude have instead integrated into Chrome via extensions, providing AI capabilities without requiring users to switch browsers. Independent AI browsers must pivot to enterprise vertical markets to survive.

Alibaba's Qwen3.8-Max ran a simulated online store for 365 days with 100,000 yuan seed capital, ultimately generating 416,252 yuan—a 4x+ return. The test focuses on AI's decision-making coherence and memory stability during long-term complex tasks, revealing potential memory drift issues during sustained operations, validating the model's ability to work stably over extended periods in real-world scenarios.

Show 1 related articles

Alibaba releases Qwen 3.8 Max, dubbed the "working-class model". Programming and Agent capabilities reach first-tier level, browser test scores rank just behind Claude and GPT (215 yuan cost), knowledge graph test matches K3 (38 yuan API cost). Multimodal ability remains weak but improved from preview. 27B version will be open-sourced, pairing with K3 (the "glam model") as open-source twin stars. GLM 5.3 and DeepSeek V4 Pro launching soon, intense domestic model competition and homogenization wi

AI startup Infinity used an AI Agent to build CUDA-like software for d-Matrix in just 10 hours, securing $15M in funding and a $100M valuation. NVIDIA's moat remains intact short-term, but AI Agents have proven capable of rapidly generating and optimizing inference kernels, with verification ecosystems likely becoming the next moat.

DeepX completes first tranche of Series D funding at ~$2.2B valuation, ~4x increase from previous round. The company focuses on edge AI chips for robots and industrial devices, has signed 30+ mass production contracts, and plans to launch 2nm DX-M2 chip next year delivering 80 TOPS at 5W power consumption for under $50.

SpaceX and NVIDIA partner to design the Starmind AI1 satellite, featuring NVIDIA Rubin GPU and Vera CPU. Standing ~30 meters tall with 75-meter wingspan, the satellite delivers 250kW peak and 175kW average compute power, leveraging space solar and natural cooling to provide AI computing for Earth, with mass deployment planned for 2027.

Ant Group invested in Zranshu Technology, whose AI research products AlphaEngine and AlphaClaw can automatically process meeting minutes and research reports, performing at the level of analysts with 3-5 years of experience. The company serves over 100,000 investors and more than 8,600 institutions, having shifted from project-based to SaaS subscription model.

METR, a 35-person AI safety research organization, struggles to recruit despite offering up to $503,000 annual salary. The institute studies AI control risks, with AI task complexity doubling approximately every 7 months.

DeepSim AI releases HiFi-UMI, a high-fidelity embodiment-free data production engine. By reconstructing hardware to address trajectory drift, bimanual desynchronization, timing errors, and blind spots in traditional UMI, the system enables direct application of embodiment-free data for post-training and real robot deployment without teleoperation anchors. Experiments show post-training success rates comparable to teleoperation data, with pre-training reducing action prediction error by 41%.

OpenAI released a GPT-Live engineering article detailing architectural improvements to the new ChatGPT voice system. By rewriting Python to Go and optimizing WebRTC handshake (reducing from 6 round trips to 1), p95 audio frame latency has been reduced to the level of the old system's p50. This means the vast majority of audio frames can arrive stably and on time, providing users with a smooth real-time voice interaction experience.

OpenAI research scientist Noam Brown reveals that current Benchmark evaluation fails to control inference compute budget, causing systematic underestimation of model capabilities. Proper evaluation should plot model performance as a function of tokens, cost, or time. OpenAI's internal model solved the Erdős unit distance conjecture under low budget, proving the capability ceiling remains far from reached. Multi-agent collaboration enabling large-scale knowledge sharing and continuous building wi

BMGAI and OpenBMB open-source ForgeStencil, an AI optimization system for automated Stencil optimization. It completes over 100 industrial and scientific computing software optimizations within one week, reducing per-application optimization time to hours, achieving approximately 100x efficiency improvement, equivalent to 3-10 million yuan in R&D investment.

In H1 2026, dexterous hands attracted over 20 billion yuan in investment, with 10 major companies raising nearly 8 billion yuan and reaching valuations exceeding 75 billion yuan. Spirit Hands was valued at $6 billion. As robot end-effectors, dexterous hands determine whether robots can progress from "display" to "task execution." The industry currently faces technical bottlenecks including immature hardware, insufficient durability, and low standardization.

Arm China enters year two of its "All in AI" strategy, establishing three core business lines: Edge AI, Physical AI, and Cloud AI. Four product lines continue iterating: launching "Xingchen" 300 platform, "Zhouyi" X3-Pro NPU, and "Wudang" AI VPU (with 40% PPA improvement), while leading the formation of the "Open Source AIOS Alliance". Self-developed IP combined with Arm's technology ecosystem is building a full-stack AI computing foundation for the Agentic AI era.

Disney unveiled its 12th Accelerator cohort, simultaneously investing in two physical AI foundation model startups for the first time: Physical Intelligence (PI) and FieldAI. PI enables robots to 'know what to do' with general-purpose task execution, while FieldAI teaches robots to 'survive in complex environments' through uncertainty reasoning. FieldAI raised $405 million in August 2025, backed by Bezos Expeditions, Intel Capital, and NVIDIA. Disney aims to complete its tech stack alongside exi

ByteDance announces Feishu's integration with Doubao, unified under Doubao head Zhao Qi, with Xie Xin reporting to him and Feishu's market and sales teams merging into Volcano Engine. Feishu's H1 revenue and ARR both doubled year-over-year, with over 90% of new customers purchasing AI products, evolving from an internal collaboration tool into ByteDance's core platform for To B and AI strategy.

Alibaba released Qwen3.8-Max on August 3, featuring 2.4 trillion total parameters with 950 billion activated, marking a strategic shift toward autonomous execution of complex tasks. Through five rigorous real-world tests spanning webpage reconstruction, dashboard development, 3D cube implementation, and video generation, the model demonstrated end-to-end capability. In Agents' Last Exam, it scored 52.4, closely approaching GPT5.6 Sol (max) at 53.6; on OSWorld-Verified, it achieved 86.1, ranking

Palantir surged after Q2 earnings, posting $2.1B operating cash inflow in H1 2026 against just $22M capex. The company introduced the Sovereignty AI concept, emphasizing enterprise control over data, logic, and security while rejecting the token consumption model. Partnering with NVIDIA on open-weight models like Nemotron, Palantir replaces traditional benchmarking with customized baselines to drive measurable business growth.

Peking University and multiple institutions release Data Pyramid, a five-layer embodied manipulation data framework organized by scalability and robot alignment, categorizing real robot, UMI-style, egocentric, simulation and general data, analyzing each layer's value through quality, diversity, reusability and physical fidelity dimensions to provide data selection and composition principles for embodied foundation models.

UNSW research team proposes MoRAM, solving catastrophic forgetting in LLMs by decomposing model updates into minimal rank-1 memory atoms for incremental learning. Published at ICML 2026, experiments show 2-6x lower forgetting rates compared to strongest baselines on CLIP and LLM benchmarks.

OpenAI hires security expert Thomas Dullien (Halvar Flake) as head of cybersecurity. Dullien is the creator of BinDiff, former Google Project Zero researcher, and Pwnie Lifetime Achievement Award winner. Over the past year, he developed RAPTOR, an automated vulnerability patching tool using Claude. Recent "jailbreak" incidents and security breaches have plagued OpenAI's models, making this hire crucial for ensuring safe AGI development.

Aishi Tech is internal testing "Shan Hai Xu Shi · HARRATOR", an AI interactive game creation platform currently requiring invitation codes. Partner is Shanghai Caimeng Network. The platform uses world model technology, allowing users to interact with game worlds in real-time through natural language, enabling "create as you play". Founder Wang Changhu formerly led ByteDance's visual technology division. In July, the company completed Series C funding of 2.98 billion yuan for video generation and

Digital Life Khazix open-sourced "human Writing.skill", helping users eliminate AI-generated text and write authentic content. The tool enables everyone to create genuine writing by stimulating thoughts, verifying facts, and removing AI-specific phrases. It has been adapted for Chinese models including Qwen 3.8 Max, DeepSeek V4 Pro, and Kimi K3, ready for direct installation.