Author: RadarAI Editorial
Editor: RadarAI Editorial
Last updated: 2026-06-19
Review status: Editorial review pending
Weekly report
周报
官方
AI热点
SpaceX completes the largest IPO in history ($2.11 trillion); Elon Musk becomes the world's first trillionaire—marking the mainstream financial adoption of the 'AI + hard tech' infrastructure flywheel.
Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.
## This Week in Summary
- SpaceX completed the largest IPO in human history ($2.11 trillion), making Elon Musk the world’s first trillionaire—marking the formal entry of the “AI + hard-tech” infrastructure flywheel into mainstream capital frameworks.
- Huawei fully shifted HarmonyOS to an Agent-native architecture, rebuilding the OS core around the principle of “intention-as-a-service.” XiaoYi has been upgraded to a system-level intelligent agent hub—cross-device, ultra-low-latency (<300ms), and fully on-device scheduling.
- Zhipu’s GLM-5.2 was fully open-sourced and benchmarked neck-and-neck with Claude Opus 4.8. With Anthropic’s strongest model now banned for U.S. export, domestic large models have crossed a critical threshold: they’re now *production-ready* for coding and office workflows.
- WeChat Pay’s “AI Dedicated Card” and Alipay’s “Abao” launched simultaneously—super apps are rapidly evolving into Agent OSes capable of closed-loop natural-language command execution. Trust and security mechanisms are now the decisive moat in gateway competition.
- Two new evaluation benchmarks—MMAE and WBench—were released, revealing fundamental capability gaps in today’s top models: audio editing (success rate <5%) and interactive video-world modeling (severe multi-turn degradation). These expose the real bottlenecks on the path to AGI deployment.
- DeepSeek secured over ¥50 billion (~$7B) in its Series A round—led by personal investment of ¥20 billion from founder Liang Wenfeng—with strategic participation from Tencent, CATL, and other industrial giants. Its non-voting governance structure underscores a deliberate, long-term commitment to R&D autonomy.
## Hot Topics
1. **SpaceX completes the largest IPO in human history—valuation: $2.11 trillion**
https://www.bestblogs.dev/article/73038fbf?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
**Core insight:** SpaceX’s listing isn’t just a milestone for commercial space—it’s the capital markets’ definitive pricing of a unified infrastructure flywheel: reusable rockets + Starlink + AI compute. It propelled Musk to trillionaire status and triggered a global recalibration of AI infrastructure financing (e.g., NVIDIA’s $20B bond issuance).
— *Actionable takeaway:* Developers should immediately study SpaceX’s “Hardware-as-a-Service (HaaS) + AI-as-Middleware” model. Try deploying a lightweight Starlink scheduler simulator locally using `llama.cpp` or `Ollama`, inspired by NASA’s open-source OpenMCT architecture—test task orchestration and agent coordination under low-bandwidth constraints.
2. **Huawei HarmonyOS fully adopts Agent architecture; XiaoYi becomes the system-level intelligent agent hub**
https://www.bestblogs.dev/article/78933caf?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
**Core insight:** HarmonyOS abandons traditional app invocation logic. Instead, it rearchitects the OS kernel around “intention-as-a-service”—enabling cross-app intent understanding, on-device agent sandboxing, and real-time decision loops (<300ms). This marks China’s first true Agent OS.
— *Actionable takeaway:* Download the HarmonyOS 7 Developer Beta and use DevEco Studio to build a “cross-device file retrieval” Agent Skill. Stress-test XiaoYi’s ability to parse and execute the full flow—WeChat document → WPS edit → Huawei Cloud sync—especially in split-screen mode. Export debug logs and feed them into a local LLM for failure root-cause analysis.
3. **Zhipu GLM-5.2 fully open-sourced—benchmarks show performance approaching Claude Opus 4.8**
https://www.bestblogs.dev/article/1c6f2bbe?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
**Core insight:** GLM-5.2 demonstrates autonomous debugging and cross-engine semantic translation (e.g., HTML → Kotlin → Minecraft rendering) on complex tasks like mechanical astronomical clocks and 3D penalty shootouts. With 1M context support and MIT licensing, it’s the first domestic model to deliver *real engineering utility*—a functional replacement for Claude.
— *Actionable takeaway:* Set up a local Zcode + GLM-5.2 environment and replicate its pipeline: “translate a web UI into Flutter code and integrate live camera stream input.” Compare API call accuracy and state management robustness against Codex outputs—and publish your quantitative report on Hugging Face Spaces.
4. WeChat Pay Launches “AI Exclusive Card”: Complete Food Delivery Search, Coupon Redemption, Ordering, and Payment Using Natural Language
https://www.bestblogs.dev/article/f30b512a?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: This product elevates WeChat Pay from a payment tool into an “intent execution platform.” Users no longer need to open Meituan or Ele.me apps—just speak a voice command to complete the entire fulfillment flow. It relies on the tight integration of three capabilities: WeChat’s AI-powered intent recognition, dynamic merchant API orchestration, and financial-grade risk control.
— Possible next step: Register on the WeChat Pay Service Provider Open Platform and call its newly released `pay.ai.invoke` API. Test end-to-end success rate and average response latency using realistic food delivery prompts (e.g., *“Order Sichuan cuisine within 3 km, rated ≥4.8, with free delivery and electronic invoice support”*). Log token consumption and categorize failure root causes.
5. Anthropic’s Latest Models — Fable 5 / Mythos 5 — Restricted Overseas by U.S. Government Ban
https://www.bestblogs.dev/article/ef9bc8e0?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: For the first time, U.S. export controls on AI have expanded beyond chips to include cutting-edge inference models. Citing “national security,” the U.S. has blocked overseas API access to models like Fable 5—accelerating global AI stack fragmentation and pushing domestic models to accelerate adoption under compliance constraints.
— Possible next step: Immediately fork Anthropic’s official SDK and replace it with compatible interfaces for GLM-5.2 or DeepSeek-V3. Run the same benchmark suite (e.g., HumanEval-X) previously used for Fable 5, then publish performance comparisons, token cost differences, and compliance statements as an open-source project on GitHub.
6. MMAE Releases First General-Purpose Audio Editing Benchmark — Top Models Achieve <5% Perfect Edit Rate
https://www.bestblogs.dev/article/29eef7eb?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: MMAE exposes a systemic weakness in current AIGC systems: poor fine-grained audio instruction following. Models struggle consistently with compound commands like *“Reduce vocal noise at second 3 while preserving background rain sounds,”* revealing fundamental gaps in multimodal alignment and temporal control.
— Possible next step: Build a minimal working pipeline on Colab’s free GPU using Whisper-v3 + AudioLDM-2, covering the full loop: speech-to-text → instruction parsing → audio editing → quality evaluation—based on MMAE’s 2,000 real-world tasks. Prioritize improving timestamp alignment and open-source the fine-tuning scripts.
7. AgentForge Platform Launches: Generate Production-Ready AI Agents in One Sentence
https://www.bestblogs.dev/article/507be283?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: Fliggy’s in-house platform lowers AI agent development to a single sentence—serving both non-technical users and Java/Node.js developers. It delivers full lifecycle support: prompt design, tool integration, memory management, and security auditing. Its launch signals maturity in industrial-grade agent tooling.
— Possible next step: Use AgentForge to build an agent that automatically compares prices for identical products across JD.com, Pinduoduo, and Taobao—and generates purchase recommendations. Export its JSON Schema configuration, reverse-engineer its tool-calling orchestration logic, then manually rebuild equivalent functionality using LangChain v0.3. Finally, compare decision consistency between the two under price-fluctuation scenarios.
8. Huawei Cloud Launches Full-Stack Agentic Infrastructure—Covering Compute, Memory, Orchestration, Security, and Industry-Specific Platforms
https://www.bestblogs.dev/article/f7b9ae97?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core Insight: Huawei Cloud is shifting beyond selling raw GPU compute. It now offers an integrated Agentic foundation—including a vector memory database (MemoryDB), a multi-Agent orchestration engine (Orca Scheduler), and an industry-specific knowledge graph sandbox (Industry KG Sandbox). This directly addresses key engineering bottlenecks enterprises face when deploying production-grade Agents.
— Possible Action: Apply for early access to Huawei Cloud’s Agentic platform. Use its MemoryDB module to build a “Medical Consultation History Memory Store,” integrate it with a locally deployed Qwen2.5-Med model, and test recall accuracy of patients’ prior allergies and medication history across five consecutive dialogues. Export and analyze the memory index structure to guide optimization.
9. Meituan’s LongCat Team Releases WBench—the First Systematic Benchmark for Interactive Video World Models
https://www.bestblogs.dev/article/53f9f508?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core Insight: WBench evaluates models using 289 real-world navigation tasks (e.g., “Find the kitchen refrigerator and open the third drawer”). It reveals a critical gap: improved video quality does *not* translate to better navigation capability—and task success rates plummet after just a few interaction rounds. This proves that purely visual representations are insufficient for grounding actions in the physical world.
— Possible Action: Select 10 high-frequency failure cases from WBench (e.g., “locate an object after opening/closing a door”), build a lightweight visual state detector using OpenCV + YOLOv10, and inject its output as a reward signal into the fine-tuning pipeline of an existing video world model. Evaluate whether explicit state awareness improves multi-turn stability.
10. Evoken Reaches ~$300M ARR—Validating Commercial Viability of AI Application Layers
https://www.bestblogs.dev/article/5b597334?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core Insight: Evoken focuses on B2B sales enablement—using AI Agents to automatically analyze customer emails and meeting notes, generate tailored proposals, predict deal probability, and recommend follow-up actions. Its $300M ARR signals a maturing phase where “application-level execution” delivers measurable ROI. Model value must now be tied to concrete business KPIs—e.g., days shaved off sales cycles or percentage-point lifts in win rates.
— Possible Action: Replicate Evoken’s publicly described “sales lead scoring logic” using local CRM data (e.g., a HubSpot-exported CSV). Train a simplified lead-priority classifier with LightGBM, compare its AUC against actual deal outcomes, and estimate potential sales team time savings if deployed.
- SpaceX completed the largest IPO in human history ($2.11 trillion), making Elon Musk the world’s first trillionaire—marking the formal entry of the “AI + hard-tech” infrastructure flywheel into mainstream capital frameworks.
- Huawei fully shifted HarmonyOS to an Agent-native architecture, rebuilding the OS core around the principle of “intention-as-a-service.” XiaoYi has been upgraded to a system-level intelligent agent hub—cross-device, ultra-low-latency (<300ms), and fully on-device scheduling.
- Zhipu’s GLM-5.2 was fully open-sourced and benchmarked neck-and-neck with Claude Opus 4.8. With Anthropic’s strongest model now banned for U.S. export, domestic large models have crossed a critical threshold: they’re now production-ready for coding and office workflows.
- WeChat Pay’s “AI Dedicated Card” and Alipay’s “Abao” launched simultaneously—super apps are rapidly evolving into Agent OSes capable of closed-loop natural-language command execution. Trust and security mechanisms are now the decisive moat in gateway competition.
- Two new evaluation benchmarks—MMAE and WBench—were released, revealing fundamental capability gaps in today’s top models: audio editing (success rate <5%) and interactive video-world modeling (severe multi-turn degradation). These expose the real bottlenecks on the path to AGI deployment.
- DeepSeek secured over ¥50 billion (~$7B) in its Series A round—led by personal investment of ¥20 billion from founder Liang Wenfeng—with strategic participation from Tencent, CATL, and other industrial giants. Its non-voting governance structure underscores a deliberate, long-term commitment to R&D autonomy.
Hot Topics
-
SpaceX completes the largest IPO in human history—valuation: $2.11 trillion
https://www.bestblogs.dev/article/73038fbf?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core insight: SpaceX’s listing isn’t just a milestone for commercial space—it’s the capital markets’ definitive pricing of a unified infrastructure flywheel: reusable rockets + Starlink + AI compute. It propelled Musk to trillionaire status and triggered a global recalibration of AI infrastructure financing (e.g., NVIDIA’s $20B bond issuance).
— Actionable takeaway: Developers should immediately study SpaceX’s “Hardware-as-a-Service (HaaS) + AI-as-Middleware” model. Try deploying a lightweight Starlink scheduler simulator locally using llama.cpp or Ollama, inspired by NASA’s open-source OpenMCT architecture—test task orchestration and agent coordination under low-bandwidth constraints.
-
Huawei HarmonyOS fully adopts Agent architecture; XiaoYi becomes the system-level intelligent agent hub
https://www.bestblogs.dev/article/78933caf?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core insight: HarmonyOS abandons traditional app invocation logic. Instead, it rearchitects the OS kernel around “intention-as-a-service”—enabling cross-app intent understanding, on-device agent sandboxing, and real-time decision loops (<300ms). This marks China’s first true Agent OS.
— Actionable takeaway: Download the HarmonyOS 7 Developer Beta and use DevEco Studio to build a “cross-device file retrieval” Agent Skill. Stress-test XiaoYi’s ability to parse and execute the full flow—WeChat document → WPS edit → Huawei Cloud sync—especially in split-screen mode. Export debug logs and feed them into a local LLM for failure root-cause analysis.
-
Zhipu GLM-5.2 fully open-sourced—benchmarks show performance approaching Claude Opus 4.8
https://www.bestblogs.dev/article/1c6f2bbe?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core insight: GLM-5.2 demonstrates autonomous debugging and cross-engine semantic translation (e.g., HTML → Kotlin → Minecraft rendering) on complex tasks like mechanical astronomical clocks and 3D penalty shootouts. With 1M context support and MIT licensing, it’s the first domestic model to deliver real engineering utility—a functional replacement for Claude.
— Actionable takeaway: Set up a local Zcode + GLM-5.2 environment and replicate its pipeline: “translate a web UI into Flutter code and integrate live camera stream input.” Compare API call accuracy and state management robustness against Codex outputs—and publish your quantitative report on Hugging Face Spaces.
-
WeChat Pay Launches “AI Exclusive Card”: Complete Food Delivery Search, Coupon Redemption, Ordering, and Payment Using Natural Language
https://www.bestblogs.dev/article/f30b512a?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: This product elevates WeChat Pay from a payment tool into an “intent execution platform.” Users no longer need to open Meituan or Ele.me apps—just speak a voice command to complete the entire fulfillment flow. It relies on the tight integration of three capabilities: WeChat’s AI-powered intent recognition, dynamic merchant API orchestration, and financial-grade risk control.
— Possible next step: Register on the WeChat Pay Service Provider Open Platform and call its newly released pay.ai.invoke API. Test end-to-end success rate and average response latency using realistic food delivery prompts (e.g., “Order Sichuan cuisine within 3 km, rated ≥4.8, with free delivery and electronic invoice support”). Log token consumption and categorize failure root causes.
-
Anthropic’s Latest Models — Fable 5 / Mythos 5 — Restricted Overseas by U.S. Government Ban
https://www.bestblogs.dev/article/ef9bc8e0?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: For the first time, U.S. export controls on AI have expanded beyond chips to include cutting-edge inference models. Citing “national security,” the U.S. has blocked overseas API access to models like Fable 5—accelerating global AI stack fragmentation and pushing domestic models to accelerate adoption under compliance constraints.
— Possible next step: Immediately fork Anthropic’s official SDK and replace it with compatible interfaces for GLM-5.2 or DeepSeek-V3. Run the same benchmark suite (e.g., HumanEval-X) previously used for Fable 5, then publish performance comparisons, token cost differences, and compliance statements as an open-source project on GitHub.
-
MMAE Releases First General-Purpose Audio Editing Benchmark — Top Models Achieve <5% Perfect Edit Rate
https://www.bestblogs.dev/article/29eef7eb?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: MMAE exposes a systemic weakness in current AIGC systems: poor fine-grained audio instruction following. Models struggle consistently with compound commands like “Reduce vocal noise at second 3 while preserving background rain sounds,” revealing fundamental gaps in multimodal alignment and temporal control.
— Possible next step: Build a minimal working pipeline on Colab’s free GPU using Whisper-v3 + AudioLDM-2, covering the full loop: speech-to-text → instruction parsing → audio editing → quality evaluation—based on MMAE’s 2,000 real-world tasks. Prioritize improving timestamp alignment and open-source the fine-tuning scripts.
-
AgentForge Platform Launches: Generate Production-Ready AI Agents in One Sentence
https://www.bestblogs.dev/article/507be283?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: Fliggy’s in-house platform lowers AI agent development to a single sentence—serving both non-technical users and Java/Node.js developers. It delivers full lifecycle support: prompt design, tool integration, memory management, and security auditing. Its launch signals maturity in industrial-grade agent tooling.
— Possible next step: Use AgentForge to build an agent that automatically compares prices for identical products across JD.com, Pinduoduo, and Taobao—and generates purchase recommendations. Export its JSON Schema configuration, reverse-engineer its tool-calling orchestration logic, then manually rebuild equivalent functionality using LangChain v0.3. Finally, compare decision consistency between the two under price-fluctuation scenarios.
-
Huawei Cloud Launches Full-Stack Agentic Infrastructure—Covering Compute, Memory, Orchestration, Security, and Industry-Specific Platforms
https://www.bestblogs.dev/article/f7b9ae97?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core Insight: Huawei Cloud is shifting beyond selling raw GPU compute. It now offers an integrated Agentic foundation—including a vector memory database (MemoryDB), a multi-Agent orchestration engine (Orca Scheduler), and an industry-specific knowledge graph sandbox (Industry KG Sandbox). This directly addresses key engineering bottlenecks enterprises face when deploying production-grade Agents.
— Possible Action: Apply for early access to Huawei Cloud’s Agentic platform. Use its MemoryDB module to build a “Medical Consultation History Memory Store,” integrate it with a locally deployed Qwen2.5-Med model, and test recall accuracy of patients’ prior allergies and medication history across five consecutive dialogues. Export and analyze the memory index structure to guide optimization.
-
Meituan’s LongCat Team Releases WBench—the First Systematic Benchmark for Interactive Video World Models
https://www.bestblogs.dev/article/53f9f508?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core Insight: WBench evaluates models using 289 real-world navigation tasks (e.g., “Find the kitchen refrigerator and open the third drawer”). It reveals a critical gap: improved video quality does not translate to better navigation capability—and task success rates plummet after just a few interaction rounds. This proves that purely visual representations are insufficient for grounding actions in the physical world.
— Possible Action: Select 10 high-frequency failure cases from WBench (e.g., “locate an object after opening/closing a door”), build a lightweight visual state detector using OpenCV + YOLOv10, and inject its output as a reward signal into the fine-tuning pipeline of an existing video world model. Evaluate whether explicit state awareness improves multi-turn stability.
-
Evoken Reaches ~$300M ARR—Validating Commercial Viability of AI Application Layers
https://www.bestblogs.dev/article/5b597334?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core Insight: Evoken focuses on B2B sales enablement—using AI Agents to automatically analyze customer emails and meeting notes, generate tailored proposals, predict deal probability, and recommend follow-up actions. Its $300M ARR signals a maturing phase where “application-level execution” delivers measurable ROI. Model value must now be tied to concrete business KPIs—e.g., days shaved off sales cycles or percentage-point lifts in win rates.
— Possible Action: Replicate Evoken’s publicly described “sales lead scoring logic” using local CRM data (e.g., a HubSpot-exported CSV). Train a simplified lead-priority classifier with LightGBM, compare its AUC against actual deal outcomes, and estimate potential sales team time savings if deployed.
← Back to Updates