SpaceX acquires Anysphere (Cursor's parent) for $60B in all-stock deal—marking a major move into dev infrastructure. DeepSeek's valuation may hit $50B, backed by Tencent and CATL; its non-voting governance structure draws industry scrutiny.
Start with the newest briefing, then Continue by task
The newest briefing gives you today's main changes in a few minutes. Use the focused routes for evidence, implementation detail, and longer analysis.
Posts
DeepSeek's Series A funding round is expected to exceed RMB 50 billion, with founder Liang Wenfeng personally contributing approximately RMB 20 billion; strategic investors—including Tencent and CATL—have jointly participated, underscoring accelerating capital concentration toward top-tier technical players in the large-model sector [2]. Meanwhile, Huawei's Xiaoyi assistant has introduced multimodal interaction upgrades—including 'Split-Screen Q&A with Xiaoyi' and 'Companion-Mode Xiaoyi'—in HarmonyOS 6.1, marking a new phase of deep software-hardware integration for domestic AI assistants [3].
Domestic LLMs accelerate adoption in code generation and office productivity: GLM-5.2 rivals Codex; Baoyu-Design enables local PPTX export. Huawei Xiaoyi upgrades multimodal interaction in HarmonyOS 6.1; Gemini 3.5 Flash embeds AI system-wide.
AI capabilities are undergoing structural leaps: recursive self-improvement, world model benchmarks (WBench/MMAE), and agentic detection are gaining traction—while power shortages and MLCC shortages constrain compute scaling [1, 3, 24].
AI infrastructure is confronting dual pressures—power shortages and critical component supply bottlenecks—while breakthroughs including a self-developed 1280-TOPS automotive-grade chip, 2nm MPW process expansion, and the world's first panel-level electrochemical deposition equipment for advanced packaging signal a multi-dimensional hardware arms race. Concurrently, the newly launched World Model Benchmark (WBench) and Multimodal Audio Editing Benchmark (MMAE) expose fundamental capability gaps in current models, notably multi-turn interaction decay and instruction execution accuracy below 5% [14][8][3].
World models and agentic infrastructure are becoming critical for LLM deployment: Huawei Cloud launches a full-stack agent platform; Kunlun Tech's TianGong AI unveils Matrix-Game 3.5, a new state-action joint-generation framework. Meanwhile, SK Hynix plans a U.S. IPO and aggressive HBM expansion amid surging global AI compute demand.
The AI industry is rapidly shifting from competition based on model capabilities to dual-track evolution—system-level agent architecture and token capital. Huawei's HarmonyOS has fully embraced the 'Intent-as-a-Service' paradigm, while Microsoft stresses that enterprises must concurrently build closed-loop ecosystems for both human capital and AI capability capital [1][2].
At HDC 2026, Huawei announced HarmonyOS's full transition to an Agent-based architecture—centered on the 'Intent-as-a-Service' paradigm—upgrading Xiaoyi from a voice assistant to a system-level AI agent.
GLM-5.2 opens fully to the public—contrasting sharply with U.S. restrictions on Anthropic's latest model—highlighting China's compliant yet accelerated LLM advancement; meanwhile, TSMC's most aggressive capacity expansion ever, plus MLCC shortages spreading to mid- and low-end specs, signals AI hardware infrastructure undergoing structural upgrades driven by both supply and demand.
AI deployment is accelerating its shift—from 'model capability' to 'systems engineering': Claude Opus 4.8's multimodal coordination, HRM-Text's hierarchical recursive reasoning architecture, and the explosive emergence of the FDE (Frontline Deployment Engineer) role collectively confirm that Harness-layer design and physical-world interface capability have become the defining technical inflection point across generations [1][3][6][11].
Physical AI is accelerating into reality—from Saic's AIVA car to Nova Fusion's FRC-SMR nuclear fusion path—marking deep AI–physical-world integration as this cycle's defining paradigm. Meanwhile, $1.8T off-balance-sheet AI infrastructure debt coexists with SpaceX's $2.11T market cap, revealing capital tensions and systemic risks behind the AI boom.
'Physical AI' is rapidly breaking through digital boundaries: Saicu Technology has launched its new brand AIVA, redefining automotive product development and human–vehicle interaction through the vision of 'AI-Defined Vehicles'—marking a deep integration of AI algorithms with real-world physical agents [1]. Meanwhile, Huawei's Pangu large model has set its sights on evolving from 'China's No. 1 → the World's No. 1', reflecting top-tier vendors' strategic elevation in response to the global large-model competition landscape [2].
SpaceX's successful IPO propelled Elon Musk to become the world's first trillionaire, with the company's opening market valuation reaching $2 trillion and its stock price surging 11% above the offering price [1].
SpaceX completed the largest IPO in human history, valued at $1.77 trillion; its self-reinforcing infrastructure flywheel—built on reusable rockets, Starlink, and AI compute—is reshaping the global commercial space industry [0][16]. Meanwhile, the co-design of Huawei's Ascend 950DT chip and DeepSeek V4 achieved a 75% reduction in inference costs, emerging as a pivotal enabler for cost-efficient, high-performance domestic large models [5].
AI agent development is becoming dramatically more accessible: Feizhu's AgentForge enables production-ready agents from a single sentence. Meanwhile, llama.cpp's hybrid inference flaws reveal gaps in foundational tooling. New CAC rules ban AI-generated negative corporate content from trending lists—shifting governance to the generative source.
OpenAI launches ChatGPT's largest-ever overhaul—transforming it from a chat tool into a unified AI platform with coding, agent capabilities, image generation, and third-party app integration.
Domestic large language models are rapidly shifting from capability competitions to commercialization breakthroughs, with programming and office productivity emerging as the most critical trillion-dollar battleground; meanwhile, next-generation reinforcement learning frameworks like Uni-Agent continue overcoming long-task stability bottlenecks, while tech giants—including Tencent and Alibaba—are pursuing fundamentally divergent ecosystem strategies for AI super-interfaces [7][12][17].
This week's AI highlights: on-device AI acceleration, multimodal long-context memory bottlenecks, and security flaws in AI agents; Cohere open-sources its first developer-focused programming MoE model; Tencent Hunyuan and Meitu/Beijing Jiaotong University achieve breakthroughs in inference operators and attribute editing frameworks.
AI-native teams are reshaping global R&D division of labor: Opendoor has disbanded its 200+ offshore team in India and shifted to a compact, U.S.-based AI-native team; meanwhile, emerging paradigms such as 3D AI Agents and Agent Harness are accelerating deployment—marking AI's transition from model-capability competition to industrialized, engineering-driven agent production. [2][3][9]
Google Cloud has officially launched the Lightning Engine for its managed Apache Spark service—leveraging vectorized native execution and optimized connectors—to deliver up to a 4.9x performance boost [1]. Meanwhile, Claude Design has been confirmed as a full-fledged Agent Harness equipped with 45 tools and 24 built-in skills, marking a new phase of engineering-grade deployment for large-model agent infrastructure [3].
On the eve of the AI application boom, knowledge graphs are emerging as the critical 'reins' to constrain large language model hallucinations and enhance Agent controllability [1]; meanwhile, the launch of Claude Fable 5 and Mythos 5 marks a leap in engineering-grade AI capabilities—real-world benchmarks include migrating 50 million lines of code in a single day and delivering a Mac app within five hours, validating their transformative potential for productivity [5][6]; concurrently, Agent economy infrastructure and localized AI hardware (e.g., Intel Arc™ Pro B70 GPU) are accelerating deployment, signaling a strategic shift from model-centric competition toward systemic, commercially viable AI ecosystems [14][11].
Anthropic is accelerating the deployment of its Mythos-tier models, launching both Claude Fable 5 and the unrestricted Mythos 5—while introducing a 'fallback to Opus 4.8' safety protocol. Meanwhile, Claude Design achieves significant token savings via its dedicated Harness architecture, but suffers from increased interaction latency [4][7][1].
AI is rapidly evolving from 'passive response' to 'proactive execution': WeChat's official Skill documentation has been released, turning millions of mini-programs into atomic, AI-callable services; Tencent Cloud has launched an enterprise-grade Agent runtime system, elevating AI from a tool to a full-fledged production system; and emerging embodied intelligence startups like KunlunXing have established operations in Yizhuang, Beijing—pursuing humanoid robot commercialization through a dual-engine strategy of 'body + brain' [9][11][8].
AI smartphones are evolving from 'answering questions' to 'executing tasks'—on-device inference capability, cross-device compute orchestration, and service-oriented protocols (e.g., MCP) have become critical differentiators. WeChat leverages its Skill Documentation to transform millions of Mini Programs into atomic, AI-callable services, accelerating the construction of an AI-era service hub [0][3].
WeChat officially launched its Skill documentation, enabling millions of mini-programs to integrate with AI services via the MCP protocol; NotebookLM upgraded to Gemini 3.5 + Antigravity, adding secure cloud computers per notebook and multi-format export + Google Search integration.