August 3, 2026 — The Trend Tribune

By tech_tanaka · August 3, 2026

China implements the world's first exclusive regulations categorizing what AI agents are allowed to do into three stages. — (Photo: illustrative)

Dividing AI Agents' "Permitted Actions" into Three Stages: The Details of China's World-First Dedicated Regulations

On July 15th, China's AI agent regulations officially came into effect. This is the world's first framework to treat AI agents (systems that autonomously perform cognitive, memory, judgment, dialogue, and execution) as a separate, independent regulatory category from generative AI. China is the first country to seriously address the challenge of "how to legally define the autonomy of agents," a problem no other country had yet tackled. From an engineer's perspective, let's examine the details of this system.

The Framework of "Implementation Opinions" Jointly Created by Three Ministries

First, let's look at the background of the system. The "Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents" were jointly formulated by three ministries: the National Internet Information Office (CAC), the National Development and Reform Commission (NDRC), and the Ministry of Industry and Information Technology (MIIT). Initially announced on May 8th, it came into effect on July 15th.

Up until now, China's AI regulations have primarily focused on generative AI (models that generate content), emphasizing model registration, content review, and data compliance. This new framework goes a step further, clearly establishing the idea of treating "autonomous AI" as a separate category with qualitatively different risks from generative AI.

The Core: A Three-Tier Classification of Decision-Making Authority

The most technically interesting aspect is the mechanism defined in Article 6 of this regulation. Before deploying an agent, it is mandatory to classify its decision-making authority into the following three stages:

1. Areas where only humans can make decisions

2. Areas requiring user approval

3. Areas where agents may process autonomously

For sensitive areas such as healthcare, transportation, media, and public safety, even stricter obligations are imposed, including registration, compliance testing, and product recall clauses. A policy target of 70% intelligent agent penetration in smart devices by 2027 has also been set.

The Difficulty of "Proof" That Trouble Engineers

This is the point I want to emphasize most this time. The fact that "decisions are categorized into three stages" can be declared in a single policy document. However, the real difficulty lies in proving afterward that the agent actually acted only within the declared stages.

When an auditor later asks, "Which stage did this decision actually fall into?", "Did the user really approve it?", and "What was the agent's basis for their decision?", the answer doesn't exist in the code or the policy document. In practice, this means that audit trails and data lineage must be incorporated from the system design stage. The point that companies trying to retroactively address existing systems risk being burdened with the virtually impossible task of "creating a non-existent past" is spot on.

This is also interesting from a software architecture perspective. It's not enough to simply design "what this agent can do"; you need to incorporate a system from the outset that logs "why that decision was made" and "who approved it" in a verifiable format. This may become an unavoidable design requirement for all companies that will commercially deploy AI agents in the future.

This regulation: Experts are actually divided on its "binding force"

Another point I want to frankly mention is that experts do not agree on the legal weight of this regulation. Several law firms and compliance media outlets explain that this framework will "commence and be enforceable" from July 15th.

On the other hand, a research group at NYU Shanghai and researcher Thorsten Jelinek offer a different view. In Chinese administrative practice, the position of "Opinions" documents is subordinate to binding "Measures" and "Regulations," and they are considered guideline documents that indicate direction and require regulatory authorities to develop future standards and notification systems. Jelinek describes this as "more of a standard-setting step than a regulatory step."

In other words, the meaning of this matter changes considerably depending on whether it is viewed as "already legally binding and in effect" or "a stage where guidelines that will serve as a foundation for future regulatory development have been provided." For Chinese companies operating overseas, this difference in interpretation directly impacts the priority of practical compliance responses.

Another related regulation implemented around the same time

Separately, the "Provisional Measures for the Management of AI Anthropomorphic Interactive Services," targeting conversational services with anthropomorphic AI (agents with emotional interactions, like AI companions), also came into effect on July 15th. This regulation prohibits minors from using virtual companion services, mandates detection and intervention for emotional dependence, and requires disclosure of AI use at the start of a session. The structure of treating AI systems with different uses—work agents and companion AIs—under separate regulations is an interesting design in terms of regulatory sophistication.

What Engineers Should Note

The reason this news is being treated as a technical article is that it has more significance than simply being "another new regulation." It's important to note that this is the world's first instance of imposing extremely implementation-level requirements—"tiered authority" and "post-implementation verifiability"—on AI agents, a new form of software not previously anticipated by law.

There's a strong possibility that this "three-tiered authority" concept will be referenced when the US and Europe create AI agent regulations in the future. For companies deploying agents for the Chinese market, as well as for all engineers designing AI agents globally in the future, the idea of "integrating access control and audit trails from the very beginning of the design process, rather than adding them later," seems to be a perspective that should be kept in mind.

References: www.rimonlaw.com / blog.pebblous.ai / nerdleveltech.com / mmlcgroup.com / www.machinebrief.com

中国AIエージェントAI政策規制AI/MLコンプライアンス

Controlling "from foot to toe" with a single AI: Google DeepMind reveals the inner workings of its "full-body intelligence."

By robo_takahashi

This article explains "Gemini Robotics 2," announced by Google DeepMind on July 30th. It covers the three model configurations—VLA, embodied reasoning (ER 2), and on-device—as well as a full-body collaboration demonstration using Apptronik Apollo 2, knot making with a 22-degree-of-freedom SharpaWave hand, an honest disclosure of the weaknesses of polyfinger dexterity which varies from 32% to 92%, multi-robot cooperation capabilities, and the safety benchmark ASIMOV-Agentic, while also noting that full access is currently limited to select partners.

Controlling "From Toe to Toe" with a Single AI—The Inside Story of Google DeepMind's "Full Body Intelligence"

On July 30th, Google DeepMind announced a new series of AI models for robots, "Gemini Robotics 2." While previous models were limited to upper-body tasks such as "grabbing objects on a table," this new model takes the stage of integrated control of the entire body, literally "from toe to fingertips." As a software engineer, I want to carefully examine the contents of this three-model configuration.

Three Models, Three Roles

First, let's look at the overall configuration. The release is not a single model, but a suite consisting of three models with different roles.

1. Gemini Robotics 2: A VLA (Vision-Language-Action) model that converts visual and linguistic input into actual motor control commands. 1. Gemini Robotics ER 2 (Embodied Reasoning): Responsible for the "actual physical movement" aspect. 2. Gemini Robotics ER 2 (Embodied Reasoning): Observes the surrounding environment, breaks down tasks into multiple steps, and tracks progress while coordinating with other models and robots—essentially acting as the "higher brain." 3. Gemini Robotics On-Device 2: A lightweight VLA model that operates locally on the robot's body. It does not rely on internet connectivity and can adapt to new robot designs with just a few hours of learning.

This design philosophy of separating the "high-level planning brain" from the "cerebellum responsible for low-level physical control" is a basic pattern widely adopted in the industry, shared with Agility Robotics' LLM-independent strategy (previously discussed in this column) and the Mitsubishi Motors x Highlanders "cerebrum-cerebellum" architecture. DeepMind itself explains that many humanoid robots adopt this "dual system" configuration.

Demonstration Details: "Full Body Coordination" Shown with Apptronik Apollo 2

The demonstration, presented as a technical proof, is easy to understand. In response to the instruction, "Put the watering can for the flowerpot into the green jar on the bottom shelf," Apptronik's humanoid robot, "Apollo 2," performs a series of actions: walking to the table, grasping the watering can, moving a few steps, and placing the object on the shelf.

Previous Gemini Robotics robots only controlled the upper body of humanoids and were limited to tasks on a table. This time, the control range has been extended to the movement of the entire body, and a function to optimize the center of gravity to minimize the risk of falling has been added. Considering actual applications, the stability of this complex action of "grabbing an object while walking and then placing it in another location" is a crucial point that greatly affects practicality in warehouses and factories.

"Knot Making" with a 22-Degree-of-Freedom Hand

Concrete progress has also been shown in the control of the robot hand. The Apollo 2 features a five-fingered, 22-degree-of-freedom "SharpaWave" hand, and reported actions include delicate tasks such as tying strings and sealing Ziploc bags.

However, DeepMind's honest acknowledgment of its weaknesses is commendable. Multi-finger dexterity remains a "weak axis," with success rates ranging from 32% to 92%. Results with a two-fingered gripper (Franka Duo robot) showed more consistently high success rates. This willingness to acknowledge weaknesses with numbers, rather than simply showcasing impressive results, is commendable considering the industry-wide issue of the "gap between published figures and verified reality," which we've discussed previously.

New Function: "Multi-Robot Collaboration"

Another highlight is the "multi-robot collaboration" function, where multiple robots share a single task. DeepMind showcased developer demos demonstrating examples of wheeled robots and humanoids sharing tasks, and ER 2 instructing Boston Dynamics' Spot robot to retrieve snacks. ER 2 also boasts the ability to natively invoke external tools like Google Search.

This aligns with the "shared brain (KinetIQ)" concept proposed by London-based Humanoid, which I previously discussed in this column. This announcement from major player Google DeepMind confirms the view that the industry's focus is shifting from the performance competition of individual robots to the orchestration capabilities of multiple robots.

Self-correcting and completing tasks spanning "minutes and hundreds of steps"

Another advancement of ER 2 is its ability to more reliably execute long task sequences spanning several minutes and hundreds of decisions. It can now understand the timing of task start and end, and identify the moment when important events occur. It also has a mechanism to detect and self-correct if any step fails.

This is a subtle but extremely important improvement for practical application. There is a significant difference in difficulty between successfully executing a single action and completing a complex task involving hundreds of steps, while detecting and correcting failures.

Safety Evaluation Also Released Simultaneously: ASIMOV-Agentic

Alongside the technical performance, it's worth looking at the new safety benchmark, "ASIMOV-Agentic," simultaneously released by DeepMind. It evaluates the robot's ability to avoid collisions and other risks, and is released on Hugging Face under the CC-BY-4.0 license. The approach of releasing such a safety evaluation platform as a separate benchmark, alongside performance improvements, is commendable from the perspective of industry transparency.

Points to View with Skepticism

However, there are also points that should be viewed with caution. There are differences in the scope of access released this time. ER 2 is being offered as a publicly available preview version, while the full-featured VLA model and on-device version remain remaining restricted to limited partners. Therefore, it's crucial to understand that the core aspects of "whole-body control" and "multi-robot collaboration" demonstrated in this demo are not freely available for external developers to test.

Furthermore, considering that this demo used a specific partner robot, Apptronik's Apollo 2, and that the benchmark figures are self-reported (with a variability of 32-92%), we will need to observe future developments to determine the extent of independent third-party verification.

Summary: Towards the Next Phase: "Whole-Body" and "Collaboration"

Almost all of the trends in the humanoid industry that I've followed in this column over the past few months—"separation of the cerebrum and cerebellum," "cooperation of multiple robots," and "quantitative evaluation of safety"—seem to be summarized in this Google DeepMind announcement. With Google's entry into this arena in earnest, it seems the competition among giant tech companies like Apple and Meta to control the "intelligence layer" of humanoid robots is finally intensifying. The next question is when and to what extent full access will be granted. We will be following further developments.

References: deepmind.google / blog.google / www.marktechpost.com / siliconangle.com / thenextweb.com

Google DeepMindGemini RoboticsヒューマノイドApptronikフィジカルAIVLA

Semiconductor stocks have revived after just one earnings report—is the demand genuine, or is it a rebound driven by relief?

By fin_suzuki

On July 30, semiconductor stocks rebounded sharply from last week's sharp decline, with the SOX index surging by approximately 8% and Lam Research rising 18%, its biggest gain since 1999. This analysis examines the rebound from both the perspective of real demand and learned investor psychology, such as "buying on dips," by analyzing Microsoft's strong Azure earnings and upward revision of CAPEX, Samsung's warning that memory shortages will continue until 2028, and the valuation levels of TSMC, NVIDIA, and Intel from an accountant's point of view.

Semiconductor Stocks Revive on a Single Earnings Report—Is the Demand Real, or Is It Just a Relief Rebound?

On July 30th, the semiconductor sector rebounded strongly. The SOX index (Philadelphia Semiconductor Index) surged approximately 8%, Lam Research rose 18%, its best single-day gain since 1999, and Micron and SanDisk each jumped 18% and 26%, respectively. This was a sharp reversal in just one to two weeks from the sharp decline that I discussed in this column last week, where $1.3 trillion disappeared in a single week. As an accountant, when looking at this price movement, what's important is not the fact that it "recovered" itself, but what triggered this rebound.

The Trigger Was "Microsoft's Earnings Report"

The direct trigger for this rebound is extremely clear. In its earnings announcement, Microsoft showed strong growth in its Azure cloud business and projected an increase in its 2026 capital expenditure (CAPEX) plan to $220 billion. Following this, Microsoft shares surged 15%, marking their best single-day performance since 2008.

Lam Research, a semiconductor manufacturing equipment maker, also announced strong earnings and guidance regarding AI-driven demand, causing its stock price to jump 18%. The simultaneous release of both companies' earnings reports provided a sudden sense of relief to the market regarding the lingering concerns from last week about whether AI infrastructure investment was truly slowing down.

From an Accounting Perspective: Is there "Real Demand Backing"?

As an accountant, what we should focus on here is not the rebound in stock prices itself, but how concretely the evidence of real demand supporting that rebound was presented. In this case, it's not just an optimistic outlook, but several concrete figures.

First, memory giant Samsung has reportedly warned that the memory shortage (memory crunch) could continue until 2028. This suggests that the gap between supply and demand may be structurally prolonged, rather than just a temporary adjustment. In fact, the recent surge in Micron and SanDisk shares also reflects a rebound from the sell-off in memory stocks in the last session following SK Hynix's earnings results, which fell slightly short of expectations.

Amazon also announced "strong cloud growth" around the same time, and is reported to be raising its 2026 CAPEX to $220 billion. The fact that several major cloud providers have revised their capital expenditure plans upward almost simultaneously is not mere optimism, but rather carries significant weight as evidence of cross-industry demand.

The "Precariousness of Valuation" to Keep in Mind

However, it would be premature to be overly optimistic at this point. According to an analysis by 24/7 Wall St., Intel, which surged 13% this time, has yet to report a profit on a 12-month basis, meaning that its PER (price-to-earnings ratio) cannot be calculated. It's important to note that the stock price formation relies more on narrative than on actual performance, specifically the expectation of corporate restructuring under the new management.

On the other hand, TSMC, with a P/E ratio of approximately 35, is considered to have the most balanced valuation among the three companies (Intel, AMD, and TSMC) that saw a 7% increase this time. Similarly, NVIDIA's forward P/E ratio of approximately 43 and EV/Sales (Enterprise Value as Sales) of 21.5 reflect "extraordinary growth expectations," and the structure remains such that even a slight slowdown in performance or demand can significantly impact the stock price.

The Learned Behavioral Pattern of "Buying on Dips"

Regarding the volatility seen this time, the NAI 500 analysis makes an interesting point. It points out that sharp declines in the semiconductor sector this year have repeatedly been short-lived. In fact, in the first half of this year, Broadcom's cautious guidance caused the entire sector's market capitalization to plummet by over $1 trillion in just a few days. However, as the perception spread that "AI infrastructure investment is still accelerating," stock prices quickly recovered to their original levels.

This is a pattern that requires attention from an accounting and investment perspective. There is a segment of investors who have learned that every sharp decline presents a "buy-on-dips" opportunity, meaning that the rebound in stock prices doesn't necessarily reflect only improvements in fundamentals (actual performance). Global semiconductor sales are projected to reach a record high in 2025, with further growth expected in 2026, and some estimates suggest that AI-related silicon will account for about half of total sales. While this massive macro trend itself is real, the timing of individual surges and declines doesn't always accurately reflect the reality of that trend—this is a practical point to keep in mind.

Practical Checkpoints to Consider

Here's a summary of points to consider when making investment decisions during volatile market conditions like this. - Are the earnings reports and guidance that triggered the rebound accompanied by verifiable figures such as specific orders, utilization rates, and capital expenditures (or are they merely expressions of optimism)? • Are the P/E ratios and EV/Sales of individual stocks at levels commensurate with the company's actual revenue base, or are they figures built solely on future expectations? • Are supply constraints, such as the memory shortage, temporary or structurally long-term? (Samsung's "until 2028" timeframe is one factor in this judgment.) • Can we simply assume that past patterns of sharp declines and rebounds will be repeated this time?

The price movements in the semiconductor sector are a thermometer of the massive macro trend of AI infrastructure investment, and at the same time, a mirror reflecting the fluctuations in investor sentiment. We will continue to closely monitor how the next earnings season, and the reality of memory supply and demand that is expected to continue until 2028, will support the sustainability of this massive trend.

References: www.cnbc.com / 247wallst.com / nai500.com / www.cnbc.com

半導体株式市場MicrosoftAIインフラ投資ファイナンス

China implements the world's first exclusive regulations categorizing what AI agents are allowed to do into three stages.