Author: RadarAI Editorial
Editor: RadarAI Editorial
Last updated: 2026-06-05
Review status: Editorial review pending
Weekly report
周报
官方
AI热点
Anthropic tops $96.5B valuation—surpassing OpenAI—as Claude Opus 4.8 enhances dynamic subagent workflows and mid-conversation system messages for enterprise use.
Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.
## This Week in AI
- Anthropic has surpassed OpenAI to become the world’s highest-valued AI company, with a $965B valuation. Claude Opus 4.8 introduces enhanced dynamic subagent workflows and mid-conversation system messages—setting a new enterprise-grade agent systems benchmark.
- Microsoft unveiled Windows’ “Dream Machine”: the MAI model family + Surface RTX Spark Dev Box + Project Solara Agent terminal—marking Windows’ official entry into the era of agent-native operating systems.
- The world’s first green-compute, full-stack AI platform launched in Inner Mongolia—delivering token-level billing and cross-model/cross-framework low-carbon compute orchestration for the first time. AI infrastructure has now entered a new phase: *measurable, tradable, and traceable*.
- A sugar refinery released the world’s first AI agent-native power supply, “Mirror”—supporting the MCP protocol and delivering sustained 160W output. Physical endpoints have officially become programmable, schedulable edge intelligence nodes.
- VAST proposed a new world model paradigm—natively decoupling state representation from rendering. Project Eden builds a persistent, collaboratively evolving virtual physics foundation, diverging decisively from mainstream video-generation approaches.
- Anthropic secretly filed its S-1 IPO draft; OpenAI announced its entry into robotics; Tsinghua University’s UniLab achieved “Mac-native, minute-scale” humanoid robot training—signaling a broad industry shift toward dual-track advancement: physical-world integration *and* capitalization.
## Top Stories
1. **Anthropic’s valuation surges to $965B—officially overtaking OpenAI**
https://www.bestblogs.dev/status/2060949916256460894?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
**Core insight**: This valuation reflects long-term market confidence in Anthropic’s enterprise-grade logical reasoning, controllable agent workflows (e.g., Dynamic Workflows), and high-ROI commercial execution—not just parameter count or general-purpose capability. Its rise directly constrains OpenAI’s narrative around enterprise service leadership.
— **What to try**: Developers should immediately refactor existing agent workflows using the Claude API + `system_message` injection mechanism—testing real-world gains in task interruption recovery and dynamic permission switching. Product teams can quickly build a lightweight CLI Goal Router as a stopgap layer (to fill the `/goal` gap), ensuring compatibility with Cursor and Hermes Desktop.
2. **Microsoft launches Windows’ “Dream Machine”: MAI models + Surface RTX Spark Dev Box + Project Solara Agent terminal**
https://www.bestblogs.dev/article/42e93f45?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
**Core insight**: This is the first OS-level solution to tightly integrate local LLM training (MAI), consumer-grade GPU inference workstations (Spark Dev Box), and physical agent execution hardware (Solara)—transforming Windows from a runtime environment into an *agent-native OS*.
— **What to try**: Developers should download the Solara SDK, deploy a local agent supporting the MCP protocol (e.g., powered by KeepThinking memory engine) on the Spark Dev Box, and connect it to the Mirror power supply—to validate end-to-end flow: *instruction → power scheduling → status feedback*. Product teams should begin adapting to the Windows Agent plugin spec—starting by packaging existing CLI tools as `.msix` apps and registering them in the MAI Tool Registry.
3. China’s First Green Computing Full-Stack AI Platform Launches—Supports Token-Level Transaction Settlement
https://www.bestblogs.dev/article/df66abe4?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: For the first time, this platform refines compute resource granularity—from “GPU hours” down to individual tokens—unifying compute orchestration, model invocation, and token-level billing into a single stack. This makes enterprise AI spending fully attributable, auditable, and optimizable—ending the era of opaque, “black-box” cost accounting.
— Practical implications: Enterprise architects should integrate the platform’s API and use its token-cost data to redesign existing Prompt Engineering SOPs—for example, correlating `Reasoning Max` invocation frequency with token consumption curves to draft a *White Paper on High-Value Reasoning Thresholds*. Developers can leverage the platform’s `token-cost-simulator` CLI tool to run cost stress tests on multi-task workflows in Cursor.
4. Sugar Refinery Unveils World’s First AI Agent–Native Power Supply, “Mirror”—Supports MCP Protocol & 160W Sustained Output
https://www.bestblogs.dev/article/df66abe4?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: Mirror transforms conventional power hardware into an edge intelligence agent with a native MCP (Model Context Protocol) interface—enabling Claude, Codex, and other AI agents to directly invoke operations like power scheduling and multi-port load balancing. It establishes the first standard for “programmable physical-world nodes.”
— Practical implications: Hardware founders should immediately fork Mirror’s open-source firmware (see its GitHub repo) and port the MCP Client module to ESP32-S3—achieving a minimal working loop: *voice command → adjust USB-C output power → return real-time thermal data*. SaaS products can embed Mirror into operations dashboards, enabling natural-language commands (e.g., *“Prioritize power delivery to three Macs in the Design Department”*) to trigger automatic load redistribution.
5. VAST Introduces a New World Model Architecture: Complete Decoupling of State and Rendering—Project Eden Enables Persistent Environments & Real-Time Multi-User Interaction
https://www.bestblogs.dev/article/2b685fe4?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: This architecture fully separates the underlying world state (physics, objects, relationships) from visual rendering—so one shared state can drive diverse frontends (VR, AR, CLI) and support cross-session state persistence and real-time collaborative editing. It lays a new foundation for digital twins and industrial simulation.
— Practical implications: Industrial software engineers should build a “production-line equipment topology graph” state service in Python using VAST’s open-source state engine (GitHub: vast-labs/eden-state-core), then expose it via FastAPI REST endpoints for Claude Code to call. EdTech products can embed Project Eden into Unity scenes—letting students modify the *state* via natural language (e.g., *“Deactivate robotic arm A3”*) and instantly see corresponding rendering changes.
6. Claude Opus 4.8 introduces mid-conversation system messages — now compatible with Prompt Caching
https://www.bestblogs.dev/status/2060487431917588680?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: This feature lets you dynamically inject system instructions *mid-dialogue* — e.g., adjust permissions, reset context, or toggle tools — without invalidating previously cached prompts. It significantly boosts control and engineering resilience for long-running Agent tasks, serving as a critical engineering upgrade on the path from Copilot to fully autonomous Agents.
— Next steps: Developers should immediately enable `--enable-system-messages` in Claude Code and test real-time agent behavior convergence using a streaming `curl -X POST` request that includes a `"system"` field. Product teams can build a “Security Sandbox Toggle” button that, when clicked, injects `{"system": "disable_tool: shell_exec"}` to instantly downgrade privileges.
7. Hermes Desktop GUI client officially launched — NousResearch delivers a ready-to-run Agent graphical interface
https://www.bestblogs.dev/status/2061851653095985399?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: Hermes is the first truly browser-free, native Agent GUI client for macOS and Windows. It supports local model loading, MCP tool registration, and visual workflow orchestration — marking a definitive shift in Agent interaction from CLI/TUI toward mainstream GUI adoption.
— Next steps: Frontend developers should download Hermes Desktop, import the `baoyu-image-gen Skill` (see Briefing #14), and use drag-and-drop to build a GUI workflow like: *User uploads PDF → Claude parses it → Codex generates charts → Local preview*. Product teams must wrap existing web tools into Hermes plugins within 72 hours (using the official Scaffold template) to secure early access to the desktop Agent ecosystem.
8. Qwen3.7-Plus: A new multimodal agent foundation model — enables one-click replication of professional desktop software workflows
https://www.bestblogs.dev/article/abb94d70?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: For the first time, this model unifies visual understanding, code generation, and tool calling in a single pipeline. It can directly interpret Figma/Sketch design files and output runnable Electron apps — collapsing the “design → development → delivery” cycle and bringing professional software replication into the minute-long era.
— Next steps: Independent developers should use the Qwen3.7-Plus API + `qwen-vl-toolkit` to input an Axure prototype and generate a complete, ready-to-run React + Tailwind project ZIP. SaaS products can embed this capability into customer self-service dashboards — letting non-technical users upload competitor screenshots and automatically receive feature comparison matrices and migration roadmaps.
9. Memory Sidecar v3.1.0 Open-Sourced: Adding Hot/Warm/Cold Three-Tier Long-Term Memory to Any AI Agent
https://www.bestblogs.dev/article/0191c665?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: This engine tackles cross-session memory fragmentation using a three-tier architecture—RAM (hot), SQLite (warm), and vector database (cold). It supports semantic search, cognitive graph construction, and interoperability with the MCP protocol—making it foundational infrastructure for enterprise-grade, stateful AI agents.
— Practical use cases: Developers can integrate Memory Sidecar into Cursor projects, configuring `hot_ttl=300s` and `warm_db=./cursor_mem.db`, then run `memory-sidecar-cli ingest --path ./src/` to auto-index their codebase. Enterprise IT teams can deploy it as a Kubernetes StatefulSet, integrate with Okta SSO, and automatically mount personalized memory spaces for each employee.
10. Y Combinator Releases Its AI-Native Organizational Transformation Framework: “Everyone-as-Agent” + Nightly Self-Evolving “Dream Cycle”
https://www.bestblogs.dev/article/0191c665?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: YC elevates agent capabilities from “tools for engineers only” to a full-fledged “organizational operating system”—accessible by every employee.
- Anthropic has surpassed OpenAI to become the world’s highest-valued AI company, with a $965B valuation. Claude Opus 4.8 introduces enhanced dynamic subagent workflows and mid-conversation system messages—setting a new enterprise-grade agent systems benchmark.
- Microsoft unveiled Windows’ “Dream Machine”: the MAI model family + Surface RTX Spark Dev Box + Project Solara Agent terminal—marking Windows’ official entry into the era of agent-native operating systems.
- The world’s first green-compute, full-stack AI platform launched in Inner Mongolia—delivering token-level billing and cross-model/cross-framework low-carbon compute orchestration for the first time. AI infrastructure has now entered a new phase: measurable, tradable, and traceable.
- A sugar refinery released the world’s first AI agent-native power supply, “Mirror”—supporting the MCP protocol and delivering sustained 160W output. Physical endpoints have officially become programmable, schedulable edge intelligence nodes.
- VAST proposed a new world model paradigm—natively decoupling state representation from rendering. Project Eden builds a persistent, collaboratively evolving virtual physics foundation, diverging decisively from mainstream video-generation approaches.
- Anthropic secretly filed its S-1 IPO draft; OpenAI announced its entry into robotics; Tsinghua University’s UniLab achieved “Mac-native, minute-scale” humanoid robot training—signaling a broad industry shift toward dual-track advancement: physical-world integration and capitalization.
Top Stories
-
Anthropic’s valuation surges to $965B—officially overtaking OpenAI
https://www.bestblogs.dev/status/2060949916256460894?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core insight: This valuation reflects long-term market confidence in Anthropic’s enterprise-grade logical reasoning, controllable agent workflows (e.g., Dynamic Workflows), and high-ROI commercial execution—not just parameter count or general-purpose capability. Its rise directly constrains OpenAI’s narrative around enterprise service leadership.
— What to try: Developers should immediately refactor existing agent workflows using the Claude API + system_message injection mechanism—testing real-world gains in task interruption recovery and dynamic permission switching. Product teams can quickly build a lightweight CLI Goal Router as a stopgap layer (to fill the /goal gap), ensuring compatibility with Cursor and Hermes Desktop.
-
Microsoft launches Windows’ “Dream Machine”: MAI models + Surface RTX Spark Dev Box + Project Solara Agent terminal
https://www.bestblogs.dev/article/42e93f45?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core insight: This is the first OS-level solution to tightly integrate local LLM training (MAI), consumer-grade GPU inference workstations (Spark Dev Box), and physical agent execution hardware (Solara)—transforming Windows from a runtime environment into an agent-native OS.
— What to try: Developers should download the Solara SDK, deploy a local agent supporting the MCP protocol (e.g., powered by KeepThinking memory engine) on the Spark Dev Box, and connect it to the Mirror power supply—to validate end-to-end flow: instruction → power scheduling → status feedback. Product teams should begin adapting to the Windows Agent plugin spec—starting by packaging existing CLI tools as .msix apps and registering them in the MAI Tool Registry.
-
China’s First Green Computing Full-Stack AI Platform Launches—Supports Token-Level Transaction Settlement
https://www.bestblogs.dev/article/df66abe4?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: For the first time, this platform refines compute resource granularity—from “GPU hours” down to individual tokens—unifying compute orchestration, model invocation, and token-level billing into a single stack. This makes enterprise AI spending fully attributable, auditable, and optimizable—ending the era of opaque, “black-box” cost accounting.
— Practical implications: Enterprise architects should integrate the platform’s API and use its token-cost data to redesign existing Prompt Engineering SOPs—for example, correlating Reasoning Max invocation frequency with token consumption curves to draft a White Paper on High-Value Reasoning Thresholds. Developers can leverage the platform’s token-cost-simulator CLI tool to run cost stress tests on multi-task workflows in Cursor.
-
Sugar Refinery Unveils World’s First AI Agent–Native Power Supply, “Mirror”—Supports MCP Protocol & 160W Sustained Output
https://www.bestblogs.dev/article/df66abe4?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: Mirror transforms conventional power hardware into an edge intelligence agent with a native MCP (Model Context Protocol) interface—enabling Claude, Codex, and other AI agents to directly invoke operations like power scheduling and multi-port load balancing. It establishes the first standard for “programmable physical-world nodes.”
— Practical implications: Hardware founders should immediately fork Mirror’s open-source firmware (see its GitHub repo) and port the MCP Client module to ESP32-S3—achieving a minimal working loop: voice command → adjust USB-C output power → return real-time thermal data. SaaS products can embed Mirror into operations dashboards, enabling natural-language commands (e.g., “Prioritize power delivery to three Macs in the Design Department”) to trigger automatic load redistribution.
-
VAST Introduces a New World Model Architecture: Complete Decoupling of State and Rendering—Project Eden Enables Persistent Environments & Real-Time Multi-User Interaction
https://www.bestblogs.dev/article/2b685fe4?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: This architecture fully separates the underlying world state (physics, objects, relationships) from visual rendering—so one shared state can drive diverse frontends (VR, AR, CLI) and support cross-session state persistence and real-time collaborative editing. It lays a new foundation for digital twins and industrial simulation.
— Practical implications: Industrial software engineers should build a “production-line equipment topology graph” state service in Python using VAST’s open-source state engine (GitHub: vast-labs/eden-state-core), then expose it via FastAPI REST endpoints for Claude Code to call. EdTech products can embed Project Eden into Unity scenes—letting students modify the state via natural language (e.g., “Deactivate robotic arm A3”) and instantly see corresponding rendering changes.
-
Claude Opus 4.8 introduces mid-conversation system messages — now compatible with Prompt Caching
https://www.bestblogs.dev/status/2060487431917588680?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: This feature lets you dynamically inject system instructions mid-dialogue — e.g., adjust permissions, reset context, or toggle tools — without invalidating previously cached prompts. It significantly boosts control and engineering resilience for long-running Agent tasks, serving as a critical engineering upgrade on the path from Copilot to fully autonomous Agents.
— Next steps: Developers should immediately enable --enable-system-messages in Claude Code and test real-time agent behavior convergence using a streaming curl -X POST request that includes a "system" field. Product teams can build a “Security Sandbox Toggle” button that, when clicked, injects {"system": "disable_tool: shell_exec"} to instantly downgrade privileges.
-
Hermes Desktop GUI client officially launched — NousResearch delivers a ready-to-run Agent graphical interface
https://www.bestblogs.dev/status/2061851653095985399?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: Hermes is the first truly browser-free, native Agent GUI client for macOS and Windows. It supports local model loading, MCP tool registration, and visual workflow orchestration — marking a definitive shift in Agent interaction from CLI/TUI toward mainstream GUI adoption.
— Next steps: Frontend developers should download Hermes Desktop, import the baoyu-image-gen Skill (see Briefing #14), and use drag-and-drop to build a GUI workflow like: User uploads PDF → Claude parses it → Codex generates charts → Local preview. Product teams must wrap existing web tools into Hermes plugins within 72 hours (using the official Scaffold template) to secure early access to the desktop Agent ecosystem.
-
Qwen3.7-Plus: A new multimodal agent foundation model — enables one-click replication of professional desktop software workflows
https://www.bestblogs.dev/article/abb94d70?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: For the first time, this model unifies visual understanding, code generation, and tool calling in a single pipeline. It can directly interpret Figma/Sketch design files and output runnable Electron apps — collapsing the “design → development → delivery” cycle and bringing professional software replication into the minute-long era.
— Next steps: Independent developers should use the Qwen3.7-Plus API + qwen-vl-toolkit to input an Axure prototype and generate a complete, ready-to-run React + Tailwind project ZIP. SaaS products can embed this capability into customer self-service dashboards — letting non-technical users upload competitor screenshots and automatically receive feature comparison matrices and migration roadmaps.
-
Memory Sidecar v3.1.0 Open-Sourced: Adding Hot/Warm/Cold Three-Tier Long-Term Memory to Any AI Agent
https://www.bestblogs.dev/article/0191c665?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: This engine tackles cross-session memory fragmentation using a three-tier architecture—RAM (hot), SQLite (warm), and vector database (cold). It supports semantic search, cognitive graph construction, and interoperability with the MCP protocol—making it foundational infrastructure for enterprise-grade, stateful AI agents.
— Practical use cases: Developers can integrate Memory Sidecar into Cursor projects, configuring hot_ttl=300s and warm_db=./cursor_mem.db, then run memory-sidecar-cli ingest --path ./src/ to auto-index their codebase. Enterprise IT teams can deploy it as a Kubernetes StatefulSet, integrate with Okta SSO, and automatically mount personalized memory spaces for each employee.
-
Y Combinator Releases Its AI-Native Organizational Transformation Framework: “Everyone-as-Agent” + Nightly Self-Evolving “Dream Cycle”
https://www.bestblogs.dev/article/0191c665?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
Core idea: YC elevates agent capabilities from “tools for engineers only” to a full-fledged “organizational operating system”—accessible by every employee.
← Back to Updates