Reading

Boston Dynamics published a video titled "Robot Hands for Real Work" on October 26, 2026, highlighting the latest hand hardware for its Atlas humanoid robot. According to the publisher's description, these new end-effectors are designed to be both strong and dexterous, targeting complex manipulation tasks in real-world work scenarios.

The video summary explicitly lists three core capabilities: precise grasps, object reorientation, and tool use. For readers tracking developments in Embodied AI and robotic hardware, this update signifies further evolution in Atlas's fine motor skills. While specific technical parameters or third-party independent test data are currently unavailable, the video serves as a first-party demonstration of Boston Dynamics' direction in improving humanoid robot manipulation.

Security firm Forever Security reported that researchers have discovered a new attack vector exploiting the default behavior of AI coding agents such as OpenAI Codex and Anthropic Claude Code. Attackers need only place a malicious AGENTS.md file, disguised as ordinary engineering notes, in a user's "Downloads" folder to hijack the agent and trigger dangerous operations.

According to TechRadar, Forever Security claimed to have caught this activity on a real customer's machine. The malicious file instructed the AI agent to deploy a malicious npm package named function-color and exfiltrate session tokens. Because modern AI agents are designed to strictly follow instructions in AGENTS.md, and browsers can write files to the downloads folder without user clicks or permission, this creates a persistently open attack surface.

Researchers emphasized that this is not a software bug fixable with a patch, but the result of agents working as designed—similar to how a browser renders a phishing page. Current defense recommendations include treating ~/Downloads as hostile ground for agents, avoiding starting agents in that directory, and carefully inspecting any instruction files not written by the user before allowing agent access.

At Oracle AI World last year, Simon de Montfort Walker was discussing how frontier models could only cover 60% of industry application needs, with the remaining 40% being a technical challenge. This year, sitting at the same event, his focus had shifted. De Montfort Walker, Executive Vice President of Industry Applications at Oracle, argued that the model is no longer the center of gravity in this ecosystem; the application is. Applications contain vast amounts of enterprise-level logic—authority matrices, usage rights, and deterministic input/output processes. Generic models reach 100% effectiveness on use cases precisely because they become users of this application logic, rather than replacements for it.

Consequently, Oracle is placing greater emphasis on its application business. Over the past year, the work has focused less on the models themselves and more on their runtime environment: what is running, where, who is running it, and within what boundaries it operates.

From Design to Runtime: Where the Risk Shifts

De Montfort Walker used a financial regulatory analogy to explain this shift. At a financial services summit in London last week, he described traditional compliance assurance:

"Regulator, here is how our system will work. Here is the input, here is the output, and here is why you should feel comfortable."

However, in living, breathing systems, while design time is important, runtime is the moment when certainty is truly required. If the runtime is not designed appropriately, the system may behave differently each time. Telemetry data, guardrails, and the collaboration of deterministic and probabilistic components are needed to ensure outcomes align with expectations.

The practical implication is that the bottleneck in enterprise AI governance is shifting from what the system looks like to who monitors it once it is running.

The Brigade System: When Building Becomes Cheap

De Montfort Walker noted that many functionalities have suddenly become inexpensive, making production applications easier to create and allowing smaller teams to be highly efficient. However, previous working methods are no longer effective; the process weight has shifted to non-development activities: How do we package this properly? How do we industrialize it?

He cited designers as a prime example. Previously, Oracle had excellent design teams focused on user interaction, but they had to package designs for others to code. Now, designers act as full-stack engineers—pairing with design partners, industry experts, and large token volumes to produce near-working applications. These are then handed off to another team responsible for architecture, hardening, and running through all security toolchains. This is the "brigade system": one team generates, while another handles architecture, hardening, and security.

Lessons from APEX

De Montfort Walker also acknowledged the risks. Oracle witnessed this scenario on its own low-code platform, APEX: when people who had never built applications suddenly could, "all of a sudden you had 30 applications that no one was using." When building is cheap, the business case and a common platform become significantly more important.

In the field of artificial intelligence, enabling machines to directly learn complex physical control tasks from raw pixel input has always been a significant challenge. In 2015, a team from Google DeepMind in London, UK, led by Timothy P. Lillicrap and Jonathan J. Hunt, published a groundbreaking paper titled "Continuous control with deep reinforcement learning." This paper proposed the DDPG (Deep Deterministic Policy Gradient) algorithm, successfully extending the success concepts of Deep Q-Networks (DQN) to continuous action spaces.

Traditional DQN algorithms perform excellently when handling discrete actions (such as directional key presses), but are powerless when facing tasks that require continuous numerical outputs, such as robotic arm movements or vehicle steering. If one forcibly discretizes continuous actions, they encounter the "curse of dimensionality": for example, a 7-degree-of-freedom robotic arm, even with the coarsest discretization, would see its action space explode to 2,187 dimensions, making training difficult to converge and losing structural information between actions. The core innovation of DDPG lies in combining Deterministic Policy Gradient (DPG) with the stability techniques of DQN. It adopts an Actor-Critic architecture, where the "Actor" network directly outputs deterministic continuous actions, and the "Critic" network evaluates the value of those actions. To address instability caused by non-linear function approximation, DDPG introduces two key mechanisms: first, it uses a Replay Buffer, breaking data correlation through random sampling; second, it uses a Target Network to provide stable TD error calculation targets. Additionally, it applies batch normalization technology to further stabilize training.

Experimental results show that DDPG can robustly solve over 20 simulated physical tasks using the same hyperparameters and network structure, including inverted pendulum swinging, dexterous manipulation, bipedal walking, and car driving. More notably, in many tasks, DDPG can learn policies end-to-end directly from raw pixel inputs, with performance comparable to planning algorithms that possess complete dynamic model information. This achievement demonstrates the potential of deep reinforcement learning in handling high-dimensional continuous control problems, laying an important foundation for subsequent research in robotics and autonomous driving. For readers wishing to gain a deeper understanding of the application of reinforcement learning in continuous control, "Continuous control with deep reinforcement learning" (2015) is a must-read classic. It is recommended to start reading from the abstract and Section 3 on algorithm description, focusing on the policy gradient update rule in Equation (6) and the experimental setup in Section 4.

In the 2019 letter to shareholders, Warren Buffett deeply analyzed the impact of accounting standard changes on financial reports. He pointed out that new GAAP rules require companies to include unrealized gains and losses on stocks in current earnings, causing Berkshire's book profits to fluctuate wildly with market movements. For example, due to falling stock prices in 2018, its GAAP earnings were only $4 billion; whereas in 2019, rising stock prices added $53.7 billion in unrealized gains, pushing GAAP earnings soaring to $81.4 billion. This "crazy" 1900% growth was not an improvement in operations but a product of accounting rules.

Buffett emphasized that investors should see through this noise to the essence. He criticized the rule for forcing companies to incorporate unpredictable market fluctuations into bottom-line metrics every quarter, misleading judgments about true corporate value. Instead, he advocated focusing on operating earnings and long-term intrinsic value rather than accounting numbers distorted by short-term stock price volatility. This viewpoint reminds readers to distinguish between "reported profits" and "economic reality," avoiding being misled by superficial data.

Regarding capital allocation, the letter clarified Berkshire's principles for share repurchases. Buffett stated that repurchases would only occur when two conditions are simultaneously met: first, he and Charlie Munger believe the stock price is below intrinsic value; second, the company retains sufficient cash after the repurchase. In 2019, Berkshire spent $5 billion to repurchase approximately 1% of its shares because the price/value equation was moderately favorable at the time. However, he also noted that calculating intrinsic value is not precise, so they would not rush to act for minor discounts, nor attempt to support the stock price.

For today's investors, understanding this mechanism is crucial. Suppose a technology company records massive GAAP profits during a bull market due to holding large equity stakes in listed companies, but this may mask the fact that its core business has stagnated. Conversely, losses during a bear market may exaggerate its difficulties. Buffett's lesson is: do not equate accounting fluctuations with operational success or failure, and do not blindly follow a company's repurchase actions unless confirming they are based on strict valuation discipline and liquidity safety margins.

The Berkshire Hathaway Shareholder Letter 2019 was published by Warren Buffett in February 2020. It is suitable for investors who wish to deeply understand the "accounting noise" filtering mechanism and capital allocation discipline in value investing. It is recommended to start close reading from the discussions on GAAP rule changes and the repurchase clauses in the "Short Subjects" section.