Weekly Report

## Weekly Overview - Core AI engineering metrics shift from “model capability” to “system efficiency”: Kuaishou validates that end-to-end Agent delivery compresses time-to-market by 80% (20 days → 4 days); Tencent Hunyuan 3’s official release has approached flagship-model-level performance in programming and Agent construction. - Embodied intelligence reaches an inflection point in real production lines: Zhi Jian Dynamics completes delivery of 100 embodied intelligent robots for CNC scenarios; EcoFlow X Series achieves fully autonomous aerial-ground operations across the entire agricultural workflow—marking the end of the “PPT phase.” - The battle for compute sovereignty intensifies: Anthropic launches 2nm in-house chip development; DeepSeek has been secretly building AI chips for one year; Samsung secures a Meta ASIC order exceeding ₩10 trillion—AI chips are transforming from “procurement items” into “strategic control assets.” - Regulation drives structural consolidation: Doubao, Qwen, and Alibaba fully deprecate AI companion services and Claude Code, effective July 15 under the *Interim Measures for Managing AI Human-Like Interactive Services*; relational agents recede, while B2B operations, industrial, and financial applications with strong use-case alignment emerge as the new primary battlegrounds. - Edge AI accelerates deployment: Apple’s MacBook Ultra roadmap explicitly integrates M6/M7 chips and OLED touchscreens; Fable 5’s discontinuation accelerates local deployment timelines; industry consensus forecasts native execution of top-tier large models (e.g., Fable 5–class) on MacBooks by 2028. - AI traffic officially surpasses human traffic: Cloudflare data shows machine requests now account for 51.3%; AI training crawlers, intelligent Agents, and automated botnets have become the new internet infrastructure—requiring infrastructure to be rebuilt “AI-native.” ## Hot Topics List 1. Kuaishou proposes the AI Native Three-Layer Restructuring Framework: Information–Process–Organization https://www.bestblogs.dev/article/ffecc3 Essence: First systematic demonstration that “individual productivity gains ≠ organizational productivity gains”; using live-stream gifting workflows as empirical evidence, it proves that optimizing isolated tools cannot unlock Agent value—information distribution mechanisms, business process nodes, and cross-functional collaboration rules must all be restructured in tandem. — Actionable: Immediately apply this framework to diagnose one core business flow in your team (e.g., customer complaint handling), annotating current bottlenecks at each layer; build a minimal closed-loop prototype using LangChain + BrowserSkill to verify whether the process layer truly shortens response time by 50%. 2. Skywork releases Skywork Tags, issuing “digital ID badges” for AI Agents https://www.bestblogs.dev/article/12 Essence: Breaks beyond traditional Bot integration models—enabling Agents to join Feishu/Slack workspaces directly, automatically inheriting contextual permissions, sharing conversation history, and participating in multi-turn collaborative evolution—thus institutionalizing organization-level Agent collaboration infrastructure. — Actionable: Create a test group on the Feishu Open Platform and integrate the Skywork Tags Demo; orchestrate two distinct Agents (e.g., meeting minutes + travel reimbursement) via Skywork Tags to complete one cross-departmental collaboration task; record context-passing accuracy rate and manual intervention count. 3. Zhi Jian Dynamics completes delivery of 100 embodied intelligent robots to production lines https://www.bestblogs.dev/article/c798db84 Essence: The world’s first large-scale deployment of embodied intelligence into a real industrial closed loop (CNC machining), validating full-stack capability across “model + data + embodiment,” and proving embodied intelligence has moved beyond lab demos to commercially viable, order-driven delivery. — Actionable: Contact Zhi Jian Dynamics to obtain production-line API documentation (via official website or WeChat public account); simulate a CNC unexpected shutdown scenario using its API; invoke a LangChain Agent to perform fault identification → retrieve maintenance manuals → generate SOP prompts → push alerts to engineers’ Feishu—all within 90 seconds. 4. Alibaba Cloud launches AgentTeams and AgentLoop: an enterprise-grade intelligent agent collaboration governance platform https://www.bestblogs.dev/article/5 Essence: The first industrialized platform supporting Agent version management (“Shepherd”), resource quotas, invocation traceability, and SLA monitoring—solving the three major enterprise pain points of large-scale Agent deployment: “invisibility,” “uncontrollability,” and “unmanageability.” — Actionable: Register on Alibaba Cloud Bailian Platform and import your existing LangChain project; use AgentLoop to create three skill-specific Agents (e.g., data query, report generation, email sending), configure dependency relationships and failure-retry policies, then export and archive a screenshot of the monitoring dashboard. 5. Anthropic initiates R&D of a 2nm AI inference chip https://www.bestblogs.dev/article/2ed62893 Essence: Collaborating with Samsung on advanced process nodes—not merely to reduce cost, but to break free from NVIDIA CUDA ecosystem lock-in and establish software-hardware co-designed compute sovereignty—signaling that AI infrastructure competition has entered the “chip sovereignty” era. — Actionable: Assess your current inference load (e.g., LLM API call volume); benchmark the same model under NVIDIA Triton and AMD ROCm, comparing throughput, latency, and cost; if GPU costs exceed 35%, initiate domestic chip adaptation planning (e.g., Ascend CANN or Cambricon MLU). 6. Doubao and Qwen announce deprecation of intelligent agent features effective July 15, 2024 https://www.bestblogs.dev/article/9029f565 Essence: Direct compliance with the *Interim Measures for Managing AI Human-Like Interactive Services*, terminating all C2C agent services involving “emotional companionship” or “role-playing,” forcing the industry to pivot from anthropomorphic hype toward high-demand B2B scenarios (e.g., operations, customer service, finance). — Actionable: Immediately audit all modules in your product containing personality definitions, anthropomorphic language, or non-task-oriented interactions; replace your current model with Tencent Hunyuan 3 and refactor into pure task-oriented Agents (e.g., “retrieve contract penalty clause” instead of “Dear assistant, please help me check…”); complete gray-release before July 10. 7. Open-source tool pxpipe significantly reduces end-to-end cost for Fable 5 https://www.bestblogs.dev/article/0049e6e1 Essence: Achieves 59%–70% cost reduction while preserving generation quality, via an innovative text→image rendering + vision-model input pipeline—the first production-ready lightweight deployment solution for multimodal Agents like Fable 5. — Actionable: Install pxpipe locally and rerun your existing Fable 5 image-generation pipeline with it; compare original vs. pxpipe solutions on a 24GB GPU for VRAM usage, per-image latency, and token consumption; output an ROI report (e.g., “$12.7 saved per 1,000 invocations”). 8. LingBot-World 2.0 supports infinite-duration editable worlds with sub-second generation and text-event injection https://www.bestblogs.dev/article/a94dd2e0 Essence: Built on a causal autoregressive architecture, it breaks traditional simulation-world duration limits—enabling real-time textual commands to dynamically rewrite world state (e.g., “earthquake strikes suddenly,” “NPC switches allegiance”)—providing a truly evolvable virtual foundation for gaming, education, and simulation training. — Actionable: Use the LingBot-World 2.0 API to create a 5-minute teaching sandbox (e.g., “cell division process”) embedding three controllable event triggers (e.g., “inhibit spindle formation”); invite five students to test it and measure event-response accuracy and world-consistency retention duration. 9. Tencent open-sources BrowserSkill: delivering localized browser bridging capabilities for AI Agents https://www.bestblogs.dev/article/24 Essence: Supports logged-in session operations and human takeover, closing the final gap between AI Agents and real-world web applications—enabling Agents to execute complex frontend interactions across tabs, form filling, and clicking JS-dynamic buttons, ending “black-box browser automation.” — Actionable: Connect BrowserSkill to your company’s internal OA system (requiring authenticated session); script an Agent to autonomously complete the full “submit leave request → select approver → upload attachment → confirm submission” workflow; record success rate and error types (e.g., CAPTCHA interception, JS load timeout). 10. China Academy of Information and Communications Technology (CAICT) releases AISHPerf—the first AI Infra operations Agent evaluation benchmark https://www.bestblogs.dev/article/12 Essence: Built on nearly 10 billion real cluster-fault logs, covering 12 high-risk scenarios including memory leaks, network partitions, and GPU driver crashes—delivering the first reproducible, quantifiable metric for Agent-based fault-resolution capability. — Actionable: Download the open-source AISHPerf test suite and run all 12 fault cases on your operations Agent; measure average mean time to repair (MTTR), root-cause localization accuracy, and secondary-failure rate; if MTTR > 180 seconds, prioritize integrating SeKV long-context inference optimization to deepen contextual understanding.