---

I. Large AI Models

Apple Enters the Ternus Era; Cook Has Not Fully Departed

Bloomberg reports that John Ternus has effectively taken over day-to-day decision-making for Apple's product lines, while Tim Cook retains roles in external representation and strategic oversight. Power is being transferred through an incumbent handover rather than a sudden leadership change. The succession by a hardware engineer signals that Apple's focus for the next decade will be on product integration rather than service expansion, with AI implementation likely appearing in edge devices and hardware forms. Markets should not expect an "Apple version of ChatGPT." For supply chains, technology route pricing and expectations regarding Apple-style AI will recalibrate alongside the transition pace.

One-Quarter of Americans Use AI Chatbots to Diagnose Symptoms

A Pew Research Center survey shows that 25% of US adults say they use AI chatbots to diagnose symptoms, with a similar proportion using them to understand medical information and do homework before treatment decisions. AI consultations have shifted from novelty to daily behavior, yet healthcare remains one of the most heavily regulated sectors. User habit formation is clearly outpacing updates to regulatory frameworks, meaning disputes over misdiagnosis liability and data privacy will only intensify. For AI healthcare application companies and internet medical platforms, these figures reassess user growth logic; for traditional medical institutions, digitalization budgets face reallocation.

OpenAI Reported to Be Bulk-Buying Tens of Thousands of Mac minis and Mac Studios for AI Training

Multiple foreign media outlets report that OpenAI has purchased tens of thousands of Apple small-form-factor computers for use in reinforcement learning and computer-use agent training environments. Building a macOS cluster outside of GPU clusters points toward a compute form factor for tasks like teaching agents to operate real computers, which is distinct from parameter-heavy large model training. Mac fleets offer low power consumption per unit and can run full desktop environments on single machines, perfectly matching the sandbox requirements for computer-use training. The secondary market for Apple devices and order structures for Mac production lines may benefit accordingly. Beyond the Apple supply chain, this heterogeneous compute combination could signal the form of next-generation training infrastructure.

• AI Is Compressing Job Replacement Cycles. A study published in Frontiers in Human Dynamics uses a three-stage framework of creation, validation, and obsolescence to analyze AI's impact on the labor market, concluding that since GPT-3 and ChatGPT, the cycle from job emergence to replacement has significantly shortened.
• US Data Center Support Polls Collapse. Common Dreams published a new poll showing a sharp decline in local support for AI data centers. Disputes over electricity prices and land usage are turning into actual political resistance, potentially further extending site selection cycles.
• Switching Between Multiple Large Models via One API. Startup 1endpoint launched a unified compatibility interface, allowing model switching without rewriting integration code, with payment based on actual token consumption. Model routing is becoming a business, where negotiation space for inference costs itself represents profit.
• Five-Step Framework BROCS for Enterprise AI Implementation. Covering Build, Run, Observe, Control, and Secure, it argues that model access is just the starting point, while observation and governance are the true challenges of enterprise-level AI.
• Voice-Only AI Companion Hers Sparks Discussion. The product persona is self-selected by characters and unchangeable by users, featuring voice but no visual avatar, pushing the boundaries of AI companion products another step further.

---

II. AI Software and Agents

AI Agent Misreads Risk Signals; After $1.2 Million Error Trade, They Built a Control Layer

The incident described by the Runplane team is the product itself. An AI trading agent misread risk signals and executed an erroneous trade worth $1.2 million. The team subsequently built a control layer sandwiched between the agent and execution systems, changing from direct agent execution to agent proposal followed by control layer approval. This middleware demand is inevitable as agents enter real trading systems, making it worth monitoring the financing pace in this niche sector over the next year.

• Trust Only Verifiable Outputs. Microsoft's developer blog released an agentic AI verification framework, with the core principle being that every agent output must be independently verifiable. This requirement is becoming standard practice for launching enterprise-grade agents.
• 90-Day Attack Record on AI Infrastructure. Wiz used honeypots to capture active attacks targeting LiteLLM, MCP servers, and mainstream AI frameworks. Techniques included remote code execution, blind prompt injection, and memory credential theft. The security debt in the MCP ecosystem is larger than most imagine, and security middleware will spawn new windows for infrastructure startups.
• Texas Governor Freezes Flock Surveillance Camera Funding. The Verge reports that as privacy backlash against AI license plate recognition cameras escalates, Governor Abbott halted new procurements ahead of elections. Government orders for AI surveillance showed its first party-level softening, with even traditionally pro-law-enforcement Republican governors hitting the brakes.
• AI Legal Advisor Officially Called Out for Errors. Australia's Fair Work Commission publicly condemned an AI-generated legal opinion with obvious errors. Simultaneously, cases involving parties using AI litigation tools surged. Court systems are beginning to confront the cost of hallucinations, making reliability certification for legal retrieval AI a necessity.
• "No AI Fridays". A spontaneous employee movement is spreading across Silicon Valley, citing research that high-frequency LLM usage accumulates cognitive debt and reduces work engagement. Some company employees have agreed to turn off AI tools every Friday. Anti-AI sentiment is upgrading from personal habit to organizational behavior, part of the same undercurrent as the collapse in data center polls.
• Ads Enter IDEs and AI Chat Waiting Gaps. Idlen launched a self-service ad manager, monetizing developers' time waiting for model outputs as ad slots, claiming monthly earnings of €20-100. Attention economics has officially invaded programming tools.
• Semantic Invariance Test Contradish. Aimed at AI teams to check consistency of model outputs for synonymous rephrasings, evaluation is refining from "is the answer correct?" to "is it stable when asked differently?"
• End-to-End Encrypted AI Context Cards. CardGo stores personal work background as encrypted cards, reusable across all chat tools via a Chrome plugin. Context management is becoming an independent product category, and the positioning battle for private AI memory layers has begun.
• Korean Stock Market Data Connected to MCP. FinBridge opened KOSPI and KOSDAQ data interfaces, directly addressing the non-US market blind spot where large models know NVIDIA but guess about SK Hynix.
• Does AI Coding Break Flow? Some Disagree. Engineer Natasha wrote an article refuting popular claims, arguing that AI used with the right collaboration methods actually extends deep work time. The debate focus shifts from tool choice to usage methods, entering the second round of the narrative on enterprise AI efficiency.

---

III. Humanoid Robots and Embodied Intelligence

Chinese Robots Are Running Faster

The Verge's "The Stepback" column dissects the China-US humanoid robot competition, highlighting dual progress in arenas and factories. Beyond stage narratives like falling down or backflips, real workstations such as depalletizing are becoming new benchmarks. The narrative of the gap between China and the US in physical AI is being rewritten, with dividends from supply chain maturity beginning to materialize in complete machines.

• Meta Data Center Robot Details Revealed. Wired broke the story, followed by Ars Technica, reporting that Meta's tested robots can plug/unplug cables and restart servers, aiming to reduce human entry into ever-expanding data centers. The investment implication is direct: operations automation rewrites the data center operating cost curve, adding a downside option to Meta's capital expenditure narrative.
• UBTECH Walks Through the Narrow Gate. TMTPost reviewed UBTECH's interim report, noting industry attention shifting from exhibition stages to factory floors. As the hype bonus fades, the ability to make money in the real physical world becomes the filter for players.
• Geekplus's Interim Report Can't Sustain the Robot Story. Released on the same day, the interim report showed H1 revenue of RMB 1.28 billion, up 25.3% YoY, and new signed orders of RMB 2.38 billion, up 35.5% YoY, while net losses expanded from RMB 50 million. The books of the warehouse robot leader demonstrate that there is still a considerable distance between the humanoid narrative and cash flow; order growth cannot mask expanding losses.
• San Jose Barber Teaches Robot to Cut Hair. abc7 reports that old-school barbers are collaborating with Instawork Robotics, using motion demonstrations to train hair-cutting robots. The Bay Area-style implementation path has moved from the lab to the barber chair. Fine motor tasks require far higher force control and generalization capabilities than transportation, making such vertical demos a shortcut for embodied intelligence to win orders.

---

IV. Autonomous Driving

• Hidden Labor Costs of Robotaxis Exposed. TechCrunch Mobility investigated remote safety operator scheduling and mental stress, revealing that bills touting full autonomy always include wages for humans. Unit economic models need recalculation based on true labor costs.
• Hacking Attack Surface of Driverless Cars. A Bloomberg interview with authors discussed autonomous driving cybersecurity risks. OTA, sensor spoofing, and cloud takeover are all attack vectors. Vulnerabilities could become hidden variables before scaled deployment, with cybersecurity insurance and automotive-grade certifications possibly being the next monetization points.
• After Autonomous Driving Legislation, Do Owners Still Pay Fines?. TMTPost analyzed liability division after the Road Traffic Safety Law revision: automakers bear responsibility for accidents above L3, while drivers remain liable for L2. "Zero takeovers" advertisements will henceforth become courtroom evidence.
• A Sustainability-Focused Challenger Arrives in Europe's Micro-Car Segment. TechCrunch reports that Italy's Liux is using eco-friendly materials to counter the cost-performance offensive of Chinese micro-EVs. This adds a sample to the European local supply chain narrative, marking the first time the price war pressure from Chinese micro-EVs is hedged by an environmental materials story.

---

V. Physical AI and Industrial Implementation

SF IntraCity Turns Instant Delivery into Urban Infrastructure, with AI Dispatch as the Pivot

TMTPost reviews SF IntraCity's transformation path. Over the past decade, instant delivery relied on rider armies to build fulfillment density, subsidies to burn in user habits, and scale stories to raise the next round of funding. In H1 2026, this logic hit rigid constraints. The company elevated its positioning from "food delivery" to "urban infrastructure," opening its capacity network to all-category instant demands and moving toward an infrastructure charging model. AI dispatch and fulfillment networks have become the pivot for the second growth curve. While the overall instant retail market continues to expand, the window for burning cash to buy growth has closed. Whether algorithms can compress average fulfillment costs per order determines who stays at the table. Shifting the valuation anchor from delivery companies to infrastructure companies also provides new metrics for heavy-asset fulfillment players like JD Logistics and Meituan.

• Huawei HarmonyOS Cockpit Summer Version Released. According to IT Home, new features include proactive address memory for the navigation Agent and the ability to select route preferences while initiating navigation via voice. The "Agentification" of cockpit systems is becoming the main battleground for differentiation in China's in-car infotainment market.
• AI Situational Awareness Map for the 72 Hours Post-Disaster. The open-source project Open SAR-COP uses AI to synthesize fragmented public disaster information into an interactive common operating picture, focusing on information integration during the golden rescue window.
• Doubao Large Model 2.2 Release Delayed. White Whale Lab reports that ByteDance has postponed the release of Doubao 2.2, originally planned for August. Internally, there is a desire to use more extensive pre-training and post-training to strengthen programming, tool invocation, and Agent capabilities. This shift in release timing represents a correction of expectations for the downstream ecosystem.
• Honor Robot 'Lightning' Makes First Overseas Debut. IT Home reports that Lightning became the first robot participant in the history of the Malaysia Marathon. After consecutively breaking human records, its first overseas appearance holds greater value as a commercial demonstration than as a competitive event.

---

VI. Macro and Market Data

• Barclays Breaks Down Profit Distribution in the AI Supply Chain. For every $100 in revenue generated by model companies, approximately $35-40 flows to the three major cloud giants, making them the most stable cash flow segment in this chain. The market implications were reflected in trading activity on the same day.
• Divergence Among US Stock AI Heavyweights. At the close on 8/28, Amazon led with a gain of 3.97%, followed by Alphabet up 1.74%, Microsoft up 1.68%, Meta up 1.21%, and Palantir slightly up 0.19%; Nvidia led the decline among heavyweights, falling 4.57%, while Tesla dropped 1.71%. Software platforms continue to absorb AI budgets, while chip leaders are seeing obvious profit-taking.
• A-Shares: Computing Hardware Recedes, Application Side Shows Greater Resilience. Zhongji Innolight fell 0.90% and Eoptolink fell 2.47%, reflecting a high-level pullback in optical modules; Cambricon closed at RMB 1046.63, down slightly by 0.13%, and Hygon Information fell 0.20%; Foxconn Industrial Internet bucked the trend, closing higher with a gain of 0.28%, and Kingsoft Office rose 1.29%.
• Rising Political Risks for Data Centers. The collapse in US polling support and Texas freezing surveillance procurement are two sides of the same signal. External costs for AI infrastructure are becoming explicit, requiring a re-evaluation of regulatory risk premiums across site selection, electricity prices, and privacy. The narrative around AI infrastructure is shifting from pure computing power to public cost considerations.

---