docs: add daily news EN 2026-05-23

Author: yanglbme

src/content/en/daily/2026-05-23.md

@@ -1,74 +1,76 @@
 ---
 title: "Awesome AI Daily | 2026-05-23"
 date: "2026-05-23"
-tags: ["Agent", "LLM", "Spatial Intelligence", "AI Coding", "AI Video", "Embodied AI", "Enterprise AI"]
-summary: "Fei-Fei Li's team releases ESI-Bench for spatial intelligence; DeepSeek Code incoming with $9.7B funding; 360 launches cloud-based OpenClaw with AI coach; MovieFlow Studio produces 80-episode drama in 3 days; China telecom operators roll out Token packages; DCI technique lets AI agents use terminal commands directly; SkillsUI completes Series F."
+tags: ["DeepSeek", "OpenAI", "Zhipu AI", "Embodied AI", "ARM", "Codex", "AI Chips"]
+summary: "DeepSeek V4-Pro API permanently slashed to 1/4 price as CATL, JD, and NetEase discuss participation in $45B valuation round; OpenAI engineer releases Codex-maxxing guide; Zhipu GLM-5.1 high-speed API hits 400 tokens/s; Fei-Fei Li's team releases ESI-Bench spatial intelligence benchmark; ARM stock surges 30% entering AI chip market; former Meituan delivery head launches embodied AI startup."
 ---
 
-## 1. Fei-Fei Li's Team Releases ESI-Bench: The ImageNet for Spatial Intelligence
+## 1. DeepSeek V4-Pro API Permanently Cut to 1/4 Price, CATL in Talks for $45B Valuation Round
 
-Fei-Fei Li's team has released ESI-Bench — a new benchmark specifically designed to evaluate embodied spatial intelligence. Unlike previous spatial intelligence benchmarks that give models optimal observations, ESI-Bench transforms the observer into an actor, closing the perception-action loop. The benchmark includes 10 task categories, 29 subcategories, and 3,081 task instances, covering four core dimensions of human spatial cognition. Testing revealed that current AI models suffer from "action blindness" and metacognitive deficits — models make judgments too early without knowing whether they've gathered enough information.
+DeepSeek officially announced that the DeepSeek-V4-Pro model API price will be permanently adjusted to 1/4 of the original pricing after a promotional discount ends on May 31 (cached input at ¥0.025/million tokens, uncached input at ¥3, output at ¥6). Meanwhile, per Bloomberg and The Information, DeepSeek is advancing a ~¥70 billion funding round with a valuation of $45 billion. CATL, JD.com, and NetEase are all in talks to participate. Founder Liang Wenfeng explicitly stated the goal is pursuing AGI, not short-term monetization.
 
-> **Awesome AI View:** ESI-Bench's significance lies in revealing that the true bottleneck in spatial intelligence isn't visual perception, but active exploration strategy. This has important implications for embodied AI and robotics research — rather than stacking more vision model parameters, the focus should shift to teaching AI to "know what it doesn't know."
+> **Awesome AI View:** DeepSeek's strategy of raising capital while simultaneously cutting prices for developers is creating a flywheel effect. CATL — the world's largest EV battery maker — investing in an AI company signals that the AI race is expanding beyond models and chips down to physical infrastructure like power and energy storage.
 
-Source: [量子位 (QbitAI)](https://www.qbitai.com/2026/05/422738.html) | 2026-05-22
+Source: [QbitAI](https://www.qbitai.com/2026/05/423162.html) | 2026-05-23
 
-## 2. DeepSeek Code Is Coming: $9.7B Funding, ACM Gold Medalist Tianyi Cui to Lead
+## 2. OpenAI Codex Weekly Active Users Top 4M, Engineer Releases Codex-Maxxing Guide
 
-DeepSeek's funding has surged to 70 billion RMB (~$9.7B), with management telling potential investors to prioritize breakthrough AI research over short-term monetization. Meanwhile, DeepSeek senior researcher Deli Chen personally posted a recruitment call to "build a Code Harness from scratch," and the official job board has densely posted 10+ Agent Harness-related positions. Tianyi Cui, former TSY Capital co-founder who joined DeepSeek in March, is expected to lead the Agent Harness team.
+Jason Liu, a new OpenAI Codex team member and creator of the 13k-star Instructor library, published an in-depth "Codex-maxxing" guide showing how to make Codex continuously take over work: running cross-month persistent threads, Heartbeats for scheduled tasks, @computer for web automation, and remote work even when the screen is locked. Codex weekly active users broke 4 million in late April. The Goal mode has graduated from experimental to正式版 — users just define a goal and acceptance criteria, and Codex autonomously works on it for hours to days.
 
-> **Awesome AI View:** DeepSeek Code's entry will reshape the AI coding tool market. With Claude Code, Codex, and Cursor already holding first-mover advantages, DeepSeek could become another "DeepSeek moment" in AI programming if it enters with an open-source strategy and cost-effective models.
+> **Awesome AI View:** Codex is transforming from a "coding tool" to an "AI employee." Jason's approach reveals the key path for Agent adoption: persistent work threads + local knowledge management + verification loops. When AI can "work overtime while you're off the clock," human-AI collaboration will fundamentally change.
 
-Source: [量子位 (QbitAI)](https://www.qbitai.com/2026/05/422624.html) | 2026-05-22
+Source: [QbitAI](https://www.qbitai.com/2026/05/423179.html) | 2026-05-23
 
-## 3. 360 Launches Cloud-Based OpenClaw + AI Coach: Zhou Hongyi Says Agents Are Too Hard, Too Expensive, Unsafe
+## 3. Former Meituan Delivery Head Launches Embodied AI Startup AtomBite.AI, Raises Seed Round
 
-360 has released a secure cloud version of OpenClaw, equipping agents with a "cloud office" including cloud hosts, cloud drives, and cloud browsers — agents keep running even when your device is off. The new "OpenClaw Coach" feature lets users train a custom agent in 2 minutes through natural conversation, without writing any configuration code. Zhou Hongyi summarized three pain points in current agent adoption: too hard, too expensive, and unsafe.
+Embodied AI startup AtomBite.AI (元节智能) completed a tens-of-millions RMB seed round led by InnoAngel Fund. Founder Dr. Wang Dong, former technical head of Meituan's delivery division, studied under academician Zhang Bo at Tsinghua. The company is focusing on restaurant kitchen scenarios rather than humanoid robots, with its core product being a "World Action Model (WAM)" — enabling robots to simulate action consequences before executing. Starting from food delivery packaging, the ultimate goal is building a digital twin kitchen operating system.
 
-> **Awesome AI View:** 360's approach represents an important trend — the value of agents is shifting from the framework itself to the decision-making experience accumulated during execution. Cloud-native and expert-agent specialization are becoming the practical path for agent deployment.
+> **Awesome AI View:** Embodied AI is shifting from "general-purpose humanoids" to "vertical applications." Restaurant kitchens — while not the sexiest赛道 — represent the most commercially viable landing ground: globally consistent demand, acute pain points, and verifiable business models.
 
-Source: [量子位 (QbitAI)](https://www.qbitai.com/2026/05/422811.html) | 2026-05-22
+Source: [QbitAI](https://www.qbitai.com/2026/05/423159.html) | 2026-05-23
 
-## 4. MovieFlow Studio: 80-Episode Drama in 3 Days — A New Solution for AI Video Industrialization
+## 4. Zhipu AI GLM-5.1 High-Speed API Launches at 400 tokens/s, Fastest Among Flagship Models
 
-Built by filmmakers with 20 years of industry experience, MovieFlow Studio consolidates script parsing, storyboarding, video generation, and timeline editing — previously scattered across a dozen tools — into a single workspace. Test results: million-word scripts structurally parsed in 3 minutes, up to 20 episodes produced per day, an entire 80-episode series completed in just 3 days, token consumption reduced by 70%+, and character consistency reaching 95%.
+Zhipu AI released GLM-5.1-highspeed, achieving 400 tokens/s — the fastest among current flagship model APIs. Built jointly by Zhipu's GLM team and TileRT, the system optimizes inference engine, scheduling, and infrastructure. In tests, it processed a 10,000-word document with multiple tasks in 10 seconds, making AI coding feel like real-time collaboration.
 
-> **Awesome AI View:** The bottleneck in AI video was never generation capability, but industrial production workflows. MovieFlow Studio's full-chain closed-loop approach shows that once technology crosses the "can generate" threshold, the real competitive moat lies in asset management, collaboration, and cost control.
+> **Awesome AI View:** High-speed APIs are often underrated. When flagship model quality meets high-speed inference, the Agent experience shifts from "wait-respond" to "real-time打磨," enabling product categories previously impossible due to latency — like real-time game world state changes and interactive design collaboration.
 
-Source: [量子位 (QbitAI)](https://www.qbitai.com/2026/05/422455.html) | 2026-05-22
+Source: [QbitAI](https://www.qbitai.com/2026/05/422511.html) | 2026-05-22
 
-## 5. China's Big Three Telecom Operators Roll Out Token Packages: China Telecom Leads, Mobile and China Unicom Lag
+## 5. Fei-Fei Li's Team Releases ESI-Bench: The "ImageNet" for Spatial Intelligence
 
-China Telecom, China Unicom, and China Mobile have all launched Token packages, pushing AI compute to consumers via phone bill payments. On-site investigation reveals that China Telecom's retail stores have price lists ready with promotional offers (as low as 1 RMB for 250K tokens), while frontline staff at China Mobile and China Unicom report receiving no notification — the full commercialization process still needs work.
+Fei-Fei Li's team released ESI-Bench — the first embodied spatial intelligence benchmark that turns the "observer" into an "actor," with 3,081 task instances. Key findings: current AI suffers from "action blindness" (bad action → bad viewpoint → worse action cascade) and metacognitive deficits (models judge prematurely without knowing if they've seen enough). Gemini reached 88.4% on partial occlusion tasks vs. humans at 87.4%, but in physical contact tasks GPT-5 scored only 64.2% (humans: 88.3%).
 
-> **Awesome AI View:** From selling voice to selling data to selling tokens, China's telecom operators are undergoing their third business paradigm shift. But the gap between promotional buzz and frontline execution shows that Token as a foundational AI commodity is still in its very early stages of pricing models, channel systems, and user education.
+> **Awesome AI View:** ESI-Bench reveals a critical fact: the bottleneck in spatial intelligence isn't weak vision models — it's near-zero action strategy. Models lack a "doubt mechanism" to assess information sufficiency — a fundamental challenge that no stronger visual encoder alone can solve.
 
-Source: [36Kr](https://36kr.com/p/3820131718402434) | 2026-05-22
+Source: [QbitAI](https://www.qbitai.com/2026/05/422738.html) | 2026-05-22
 
-## 6. DCI: Let AI Agents Use the Terminal Directly, Bypassing Vector Databases
+## 6. ARM Stock Surges 30%, AGI CPU Demand Doubles to $2B in 6 Weeks
 
-Researchers from multiple universities propose "Direct Corpus Interaction" (DCI), a technique that lets AI agents bypass embedding models entirely and use command-line tools like grep and find to search raw corpora directly. DCI-Agent-Lite runs on GPT-5.4 nano, while DCI-Agent-CC uses Claude Sonnet 4.6. The research points out that semantic retrieval becomes a bottleneck in multi-step tasks because retrievers "decide too early what the agent is allowed to see."
+ARM stock rose 27% in three weeks, jumping over 15% on May 21 to hit a record $259. The core driver: ARM's first-ever self-designed chip — the AGI CPU, purpose-built for AI Agent scenarios, with first customers including Meta, OpenAI, and Cerebras. Customer committed demand doubled from $1B to $2B in just 6 weeks. Bernstein initiated coverage with a $300 price target, forecasting the server CPU market will grow 4x to $137B by 2030.
 
-> **Awesome AI View:** DCI's core insight is that enterprise data — live logs, tickets, code commits — is constantly changing, not a static document collection. Vector indexes are always snapshots, but agents need real-time awareness of current state. This is a critical problem that has been overlooked for too long.
+> **Awesome AI View:** Agentic AI is rewriting data center fundamentals. As CPU:GPU ratios shift from 1:8 to 1:1, the CPU transforms from a sidekick to the "dispatch center." ARM's pivot from "selling blueprints" to "building infrastructure" is essentially a bet on the Agentic AI era.
 
-Source: [VentureBeat](https://venturebeat.com/orchestration/your-ai-agents-need-a-terminal-not-just-a-vector-database) | 2026-05-22
+Source: [GeekPark](https://www.geekpark.net/news/364681) | 2026-05-22
 
-## 7. SkillsUI Completes Series F: Enterprise AI Entry Layer with 41M Users
+## 7. OpenAI Reportedly Launches IPO This Week, Targeting September Listing
 
-Shenzhen-based Tubzhan Intelligence (兔展智能) has launched SkillsUI, a platform that transforms existing enterprise ERP, OA, and CRM systems into AI-callable Skills via API, enabling "conversation-as-action." Users simply speak their needs, and AI automatically identifies intent, routes, invokes workflows, and returns interactive cards. The platform has completed Series F funding and serves over 41 million users. Its self-developed "Tuling" model is Guangdong Province's first filed visual-spatial intelligence LLM.
+Per GeekPark's daily briefing, OpenAI has officially initiated its IPO process this week, targeting a September listing. This would be the largest IPO in AI industry history. Meanwhile, Google announced it will add more ads in AI mode, sparking debate about AI search monetization.
 
-> **Awesome AI View:** SkillsUI represents the right path for enterprise AI adoption — don't replace legacy systems, add an AI entry layer on top. The shift from "humans finding systems" to "AI dispatching systems for humans" will be the biggest paradigm change in enterprise IT architecture over the next three years.
+> **Awesome AI View:** The OpenAI IPO marks the AI industry's transition from "burning cash" to "delivering returns." Post-listing, OpenAI will face stricter profitability pressures, which could affect its AGI research investment pace.
 
-Source: [量子位 (QbitAI)](https://www.qbitai.com/2026/05/422615.html) | 2026-05-22
+Source: [GeekPark](https://www.geekpark.net/news/364611) | 2026-05-22
 
-## 8. Kore.ai Launches Artemis AI Agent Platform, Takes on Salesforce and ServiceNow
+## 8. NVIDIA Vera CPU Delivered to Anthropic, OpenAI — CPU Market Heats Up
 
-Kore.ai has released the Artemis AI agent platform, betting on neutrality and a proprietary intermediary language as its core strategy — emphasizing that AI, not human developers, should do most of the heavy lifting in agent building. The platform enters a crowded field where every major tech vendor is racing to become the default infrastructure for enterprise AI agents.
+NVIDIA VP Ian Buck delivered the first batch of Vera CPU systems to Anthropic, OpenAI, Oracle, and SpaceXAI on May 19. Built on ARM architecture and sold as a standalone product, Vera CPU's launch coincides with ARM's AGI CPU — both a GPU giant and an IP licensing giant entering the CPU market in the same month, jointly validating the thesis that "Agentic AI is reshaping CPU demand."
 
-> **Awesome AI View:** The enterprise AI agent platform war is heating up. Kore.ai's "neutrality" strategy hits a real pain point — enterprises don't want vendor lock-in. But the true moat isn't the platform itself; it's who can accumulate the most industry know-how and distill it into reusable decision-making experience.
+> **Awesome AI View:** When a GPU company starts selling CPUs, it signals a shift in data center bottlenecks. The dual entry of Vera CPU and AGI CPU previews a value re-rating for CPUs in the AI Agent era.
 
-Source: [VentureBeat](https://venturebeat.com/technology/kore-ai-launches-artemis-ai-agent-platform-expands-challenge-to-microsoft-and-salesforce) | 2026-05-22
+Source: [GeekPark](https://www.geekpark.net/news/364681) | 2026-05-22
 
-## Other Developments
+## Other Updates
 
-- **Alibaba's Qwen3.7-Max** scored 44.5 on the Apex Math Reasoning benchmark, surpassing Claude Opus-4.6 Max (34.5), supports 35 hours of autonomous operation and external harness integration. (Source: [VentureBeat](https://venturebeat.com/technology/alibabas-proprietary-qwen3-7-max-can-run-for-35-hours-autonomously-and-supports-external-harnesses-like-anthropics-claude-code) | 2026-05-22)
+- **360 "Cloud Lobster"**: 360 launched a cloud version of its Agent platform with "Cloud Office" for persistent operation, plus "Lobster Coach" for 2-minute Agent training. (Source: [QbitAI](https://www.qbitai.com/2026/05/422811.html) | 2026-05-22)
+- **DeepSeek Code Team Hiring Surge**: DeepSeek posted 10+ Agent Harness positions in one day, with senior researcher Deli Chen personally recruiting for "building Code Harness from scratch." (Source: [QbitAI](https://www.qbitai.com/2026/05/422624.html) | 2026-05-22)
+- **Musk's Wealth May Hit $1 Trillion**: The world's first trillionaire may be imminent. (Source: [GeekPark](https://www.geekpark.net/news/364611) | 2026-05-22)