ReadingResearchRadarInvestment framework
Sign in / Sign up
AllTechnologyProductsInvestingPublic marketsOther

Reading

2026-11-17100 posts

Refresh

Cloud Provider Lambda Deploys NVIDIA Photonics Switches: Self-Tests Show 4% Latency Drop; NVIDIA Claims 5x Efficiency Gain

Material
2026-11-17 00:00 GMT+8

US-based cloud provider Lambda announced on November 16, 2026, that it has deployed NVIDIA Photonics co-packaged optics (CPO) switches in its large-scale AI cluster connecting more than 10,000 NVIDIA GB300 NVL72 GPUs. This move aims to address the high energy consumption and stability issues caused by network bottlenecks in ultra-large-scale AI training.

According to data released by Lambda, compared to traditional pluggable optical modules, the CPO solution reduced network latency by approximately 4% at the rack level and up to 10% at the Pod level, while maintaining the same throughput. In terms of thermal performance, Lambda reported a median optical engine temperature of 36°C in its deployment, compared to 51°C for conventional pluggable optics. Lower temperatures help maintain signal integrity and extend mean time between interruptions (MTBI).

The core of NVIDIA Photonics technology lies in integrating optical engines directly into the switch chip package, replacing traditional pluggable transceivers. This architecture shortens electrical transmission paths and reduces the use of retimer chips. NVIDIA claims the technology offers up to 5x better power efficiency and 15x higher reliability, but these are general specification claims regarding the architecture, not independently verified results from Lambda's specific deployment.

The deployment utilizes liquid cooling to manage heat generated by high-density computing. Each switch features 144 ports running at 800Gb/s and includes 18 field-replaceable external light source modules. This design allows technicians to replace faulty light sources without swapping the entire module, further improving operational efficiency.

Sources:NVIDIA Blog

Research

Amazon Web Services: Retrieval Instructions Control How Answers Are Written, Metadata Filters Control What Is Retrieved—A Multi-Tenant Customization Approach for Shared Knowledge Bases

Material
2026-11-16 23:46 GMT+8

The same product documentation and policy files often serve different audiences: customers in three sales regions need their own languages and date formats, while internal support teams need short ticket notes. Previously, implementing this differentiation on Amazon Bedrock Managed Knowledge Base required developers to disable response generation, call models themselves, and maintain prompts in every application. An article published by the Amazon Web Services Machine Learning Blog in November 2026 introduced a new mechanism: retrieval instructions—a versioned managed resource referenced by each request, eliminating the need to duplicate maintenance across applications.

The article's core distinction is noteworthy: retrieval instructions change how answers are written but do not limit what is retrieved. Instructions have four fields—persona sets identity, tone, and language; answerFormat specifies structure, date and currency formats, and citation style; retrievalGuidance directs the agent on how to interpret questions and prioritize document searches; answerPolicy determines whether to answer and how to handle out-of-scope questions. Among these, persona, answerFormat, and answerPolicy only take effect when generateResponse is true, while retrievalGuidance applies regardless of response generation. However, the author explicitly warns that these fields are not authorization or data-isolation boundaries. To ensure different tenants see only their own documents, metadata filters or separate knowledge bases must be used to control the retrieval scope.

This distinction has an easily overlooked consequence in practice: answer policies are best-effort. The article gives an example where, even if the instruction states that restricted engineering notes should never be discussed, if the agent retrieves those notes, the answer might still mention their mechanisms or record IDs. To truly exclude restricted content from answers, one should use metadata filters to exclude documents containing that content from the retrieval scope. In other words, the instruction layer handles style control, while the filter layer constitutes the security boundary—confusing the two leads to situations where data appears protected by policy but has actually leaked.

Regarding version management, the article recommends pinning specific versions in production environments rather than using DRAFT. After updating an instruction, one must poll to wait for asynchronous review completion; creating a version or updating again during the review returns HTTP 409. When publishing changes, the application layer maintains version mappings for each tenant, enabling canary releases, promotions, or rollbacks by modifying the mapping—Amazon Bedrock itself does not route traffic. A single request can reference only one retrieval instruction, so users belonging to both a region and a team require a combined instruction covering both, or must accept a compromise of using the team instruction plus a regional tenant line. The default quota is 50 retrieval instructions per account per region.

Sources:Amazon Web Services

Research

Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization

Verified 2026-11-17 00:05 GMT+8

Long-term reading · 《Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization》(2016)

In the field of computer vision, although Convolutional Neural Networks (CNNs) perform exceptionally well, their "black box" nature makes model decisions difficult to interpret. Researchers Ramprasaath R. Selvaraju and others from the Georgia Institute of Technology in the United States proposed the Grad-CAM technique, aiming to generate "visual explanations" for CNN-based models to make them more transparent and interpretable. This method utilizes gradients flowing into the final convolutional layer for a target concept to generate coarse localization maps that highlight regions critical for predicting that concept.

The core advantage of Grad-CAM lies in its generality. Unlike previous methods, it applies to a wide family of CNN models, including those with fully connected layers (e.g., VGG), those used for structured outputs (e.g., image captioning), and those handling multimodal inputs or reinforcement learning tasks, without requiring architectural changes or retraining. The study notes that pixel-space gradient visualizations (such as Guided Backpropagation), while high-resolution, lack class discriminability; whereas CAM or Grad-CAM exhibits high class discriminability. To combine the strengths of both, the authors propose Guided Grad-CAM, which fuses pixel-space gradient visualization with Grad-CAM to produce visualizations that are both high-resolution and class-discriminative.

Furthermore, Grad-CAM is not only used for diagnosing model failure modes but also relates to establishing human trust. The paper indicates that seemingly irrational predictions often have rational explanations, and these visualizations help identify dataset biases. More importantly, human studies show that Grad-CAM explanations enable untrained users to successfully distinguish between "stronger" and "weaker" deep networks, even when both make identical predictions. This provides an important tool for building trustworthy AI systems.

"Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization" was published by Ramprasaath R. Selvaraju et al. in 2016. For readers seeking to understand internal mechanisms of deep learning models, enhance transparency, or engage in explainable AI research, this paper is a must-read classic. It is recommended to start with the abstract and introduction, focusing on Section 3 (Methodology) and Section 5 (Human Studies).

Sources:《Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization》(2016)

Research

Warren Buffett's 2025 Thanksgiving Message to Berkshire Shareholders: Luck, Legacy, and the Future of Berkshire

Verified 2026-11-17 01:07 GMT+8

Long-term reading · 《Warren Buffett's 2025 Thanksgiving Message to Berkshire Shareholders》(2025)

On November 10, 2025, Warren Buffett released one of his final public letters as Chief Executive Officer of Berkshire Hathaway. In this letter, he not only announced donations of a portion of his shares to family foundations but also deeply reflected on the role of "luck" behind personal success and the governance direction for the company's future.

First, Buffett emphasized the profound impact of randomness at birth on individual destiny. He noted that being born in 1930 in the United States, healthy, intelligent, and white male gave him enormous initial advantages. He stated: “Through dumb luck, I drew a ridiculously long straw at birth.” This acknowledgment of luck is not mere modesty; it is an exposition of the foundation of his business philosophy: in an uncertain world, recognizing and respecting uncontrollable factors is a prerequisite for rational decision-making.

Second, the letter details the logic of succession planning and wealth transfer. Buffett pointed out that his children are approaching retirement age, and to ensure they can fulfill their charitable missions while still energetic, he decided to accelerate lifetime gifts. He wrote: “To improve the probability that they will dispose of what will essentially be my entire estate before alternate trustees replace them, I need to step up the pace of lifetime gifts to their three foundations.” This action demonstrates how to convert personal intent into specific institutional arrangements to ensure the continuity of values.

Finally, Buffett maintained a clear yet optimistic outlook on Berkshire’s future. He acknowledged growth limitations imposed by the company’s size but emphasized its risk-resilience and shareholder-oriented management culture. He reminded investors: “Our stock price will move capriciously, occasionally falling 50% or so as has happened three times in 60 years under present management. Don’t despair; America will come back and so will Berkshire shares.”

“Warren Buffett's 2025 Thanksgiving Message to Berkshire Shareholders” was written by Warren Buffett in 2025. It is suitable for readers seeking to understand long-term investment mindset, corporate succession planning, and how to find certainty amid uncertainty. It is recommended to begin with the sections on “luck” and “succession planning” to grasp the core intellectual thread.

Sources:《Warren Buffett's 2025 Thanksgiving Message to Berkshire Shareholders》(2025)

Research

You’re all caught up in this view

Sign in / Sign up
ReadingResearchRadarInvestment framework

The article also demonstrates broader usage: answerFormat can demand strict JSON output for downstream code consumption, saving a separate generation step; retrievalGuidance can teach the agent domain vocabulary and preferred sources; answerPolicy can fix disclaimers and citation formats. These usages are not limited to multi-tenant scenarios, though multi-tenancy serves as the article's primary narrative framework.