Reading

US-based asset manager Franklin Templeton officially launched two new Exchange-Traded Funds (ETFs) on October 29, 2026: the Franklin U.S. Equity Bitcoin DRIP Index ETF (NASDAQ: FBEQ) and the Franklin U.S. Innovation Bitcoin DRIP Index ETF (NASDAQ: FIBT). The core innovation of these funds lies in their "Dividend Reinvestment Plan" (DRIP) strategy—they do not use stock dividends to purchase more shares but instead systematically convert them into Bitcoin exposure.

FBEQ tracks approximately 500 large-cap U.S. stocks, while FIBT focuses on 100 Nasdaq innovation companies. In their initial allocation, both maintain a 95% equity position and a 5% Bitcoin position. To control risk, the funds rebalance quarterly: if Bitcoin exposure exceeds 5%, it is reduced to 4.5%; if it is at or below 5%, the allocation remains unchanged. This structure aims to allow investors to retain traditional stock market returns while gradually accumulating digital assets through compounding effects.

However, this strategy comes with significant trade-offs. Compared to traditional dividend reinvestment, Bitcoin's high volatility means this portion of capital may face larger drawdown risks. As of early November 2026, Bitcoin prices were down approximately 38.5% from their October 2025 all-time highs, amid what is being called a "crypto winter." While the fund's allocation caps limit single-asset concentration, they cannot eliminate losses during broader market downturns. For investors with higher risk tolerance and longer horizons, this serves as an automated method to accumulate Bitcoin; for conservative investors, it adds unnecessary portfolio volatility.

According to Tom's Hardware, a new open-source project called OccupySaaS has launched with the goal of creating "clean-room" reimplementations of common commercial Software-as-a-Service (SaaS) offerings using AI tools. The project aims to provide free, open-source alternatives to Auth0 and Salesforce, named ZeroAuthy and SalesCompel respectively.

Since these commercial services run on vendor servers, direct binary decompilation is impossible. Instead, the development team guides AI code generation based on public documentation, API specifications, and open standards. The project emphasizes strict ethical boundaries: all work must be fully independent clean-room reimplementations, prohibiting the use of confidential information or source code from previous employers, as well as copying UI assets or trademarks.

Technically, the project primarily uses Go to compile dependencies into a single binary, facilitating self-hosting via Docker containers. While addressing user demand driven by subscription fatigue and price hikes, the ability of such community-driven, AI-generated complex enterprise software to achieve production-grade stability and feature parity remains to be seen.

In a blog post published by Confluent in November 2026, a judgment was presented: when AI agents stop merely helping people answer questions and instead autonomously execute business decisions, the consequences of data staleness undergo a qualitative change. In the past, outdated data or inaccurate understanding would only lead to wrong answers, which people could correct, question, or ignore. But when agents act without human oversight, outdated data no longer produces wrong answers; it produces wrong actions—issuing an incorrect credit, promising a delivery, or releasing a batch of goods. Correcting the data afterward may not be able to undo the actions that have already occurred.

The premise of this judgment is that the role of AI is changing. Confluent believes that generative AI has evolved from drafting and summarizing to writing code and executing complex tasks. Now, we tell agents what we want done and let them figure out how, including making decisions that once required a person in the loop. Humans still set the direction and boundaries, but the decisions themselves can occur at machine speed.

Confluent points out that most existing enterprise data architectures are designed to describe the past—whether the state from minutes, hours, or months ago. Now, we are asking the same architecture to help run the present. Simply getting data into a stream, lakehouse, or warehouse within seconds is not enough, because fast ingestion does not equal fast action. Information must be processed and served, along with its meaning and context, at the speed required for decision-making.

Based on this issue, Confluent named the architecture that solves it the Streamhouse and outlined three requirements. First, Always Current: build state from history, then maintain it continuously as changes occur, retaining the ability to replay, correct, and recompute. Second, Decentralized: operational and analytical systems retain their respective responsibilities, sharing and reusing state and its meaning without first consolidating onto a single physical platform. Third, Mission Critical: information driving production decisions must meet requirements for correctness, availability, security, recoverability, and traceability.

Confluent acknowledges that the underlying technologies for the Streamhouse already exist. Connectors capture operational changes, a durable event backbone makes changes processable and replayable, stateful processing continuously maintains useful representations of business state, shared semantic models connect business definitions across systems, governance and lineage provide control and traceability, and serving systems make the maintained state available in the form of streams, queries, APIs, and tables. These capabilities can be implemented using different technologies such as Kafka Connect, Apache Kafka, Apache Flink, and Apache Iceberg; the Streamhouse does not prescribe a specific technology stack.

The article lists three cases illustrating how this pattern is already forming in production. Siemens Healthineers uses Apache Flink to process and enrich operational data before it reaches the warehouse, supporting multiple downstream consumers through a shared streaming backbone, thereby eliminating expensive, multi-hour batch reprocessing cycles ahead of Snowflake. Neuron Systems uses durable event streams to coordinate AI agents providing live sports commentary, allowing Hindi, Spanish, and English commentators to react to the same play based on the same underlying game state, preventing contradictory commentary across agents. Yelp independently described its own Streamhouse architecture using AWS services, Apache Flink, and Apache Paimon, without relying on Confluent, reducing analytics data latency from approximately 18 hours to minutes and lowering storage costs by 80%.

Non-profit AI research institute Ai2 announced on November 4, 2026, that it has secured multimillion-dollar, three-year funding from Google.org and technical support from Google Research to develop NeuralGCM+, an open-source climate model. The initiative aims to combine physical equations with artificial intelligence to simulate coupled atmosphere and ocean dynamics, providing faster and more reliable long-term climate projections.

Unlike traditional climate models which are computationally expensive, or existing AI weather models that focus on short-term forecasts (days to months), NeuralGCM+ targets predictions of regional temperature and rainfall trends over decades up to mid-century. This capability is critical for water resource planning and community adaptation to climate change.

Technically, NeuralGCM+ extends Google Research's existing NeuralGCM architecture, which uses physical equations for large-scale dynamics and AI for smaller-scale processes like cloud formation and rainfall. Ai2 plans to link this with ocean and sea-ice models to capture how oceans absorb and redistribute heat—a major influence on long-term climate. The team also intends to improve performance on widely used GPUs and release documented workflows to enable outside researchers to train and adapt the models.

While Ai2's previous ACE models demonstrated that AI could speed up specific experiments by roughly 100 times, purely data-driven models struggle to reliably project future climates they have not experienced. NeuralGCM+'s hybrid design attempts to address this by giving the AI a simpler learning problem, though demonstrating reliable future projections remains a central research task.

Open-source inference engine vLLM announced on November 4, 2026, day-0 support for Moonshot AI's Kimi K3.1 Pro model. The model features 2.8 trillion total parameters (104 billion active) and a 1 million token context window, with its core architectural change being the introduction of Sparse Attention (DSA) across all 24 full-attention layers.

Key technical implementations by vLLM include: First, a packed KV cache mechanism that unifies MLA latents, indexer keys, and KDA states into contiguous blocks, supporting FP8 or NVFP4 quantization formats to significantly reduce memory footprint and improve serving density. Second, token-sharded prefill, which achieves up to 1.69x faster Time To First Token (TTFT) compared to head-sharded methods at 340K token lengths. Third, Decode Context Parallelism (DCP) combined with a sparse indexer, speeding up the attention module by 1.3–1.5x and achieving up to 375.9 tokens/s per user throughput on GB300 GPUs.

Additionally, vLLM introduced zero-mismatch Reinforcement Learning (RL) for the first time, ensuring bit-identical log-probs between trainer and rollout engines, maintaining KL divergence at zero during PPO training and enhancing RL stability. These optimizations have been validated on NVIDIA GB300/Rubin, AMD MI350 series, Google TPU, and Alibaba T-Head PPU.

Chinese tech giant Tencent Holdings Ltd. released its latest large language model, Hy4, on November 4, 2026. The model introduces "computer-use" capabilities for the first time in the Hy series, leveraging advanced visual understanding, reasoning, and tool-calling to deliver long-horizon agentic performance.

According to SiliconANGLE, Tencent stated that Hy4 marks a significant leap in performance over the Hy4 Preview released on August 28. The new model weighs in at 770 billion total parameters with 49 billion active parameters and supports a 1 million token context window, placing it in the same range as current open-weights frontier models from European and Chinese developers, including Mistral AI, Moonshot Technology, and Z.AI.

Tencent reported substantially scaling both pre-training and post-training, achieving gains in instruction following, long-context understanding, agentic long-horizon tasks, and deep research. The company also noted that the previous version suffered from excessive reasoning and over-verification during its "thinking" processes, which the new model has optimized.

In terms of practical application, Hy4 focuses heavily on real-world productivity for workplace collaboration, human experience, and professional work. Tencent highlighted improvements in web development, document summarization, visual creative tasks, and front-end/game development. For complex game logic, state management, and orchestration, Tencent said Hy4 integrates deeply with major engines like Unreal Engine, Unity, and Godot. The company claimed a 19.1 percent improvement in the GameDev bench and 9.8 percent in GameLogic benchmarks compared to the preview, although actual baseline numbers remain internal and unreleased.

Additionally, Tencent advanced Recursive Self-Improvement using the Hyra 2.0 autonomous AI research agent system to enhance R&D capabilities. Hy4 is currently available on Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom’s AI Assistant, and via third-party sources like OpenRouter. Pricing remains low, with input tokens at approximately $0.90 per million, output tokens at $2.70 per million, and cached input tokens at $0.045 per million. A Flash version is scheduled for release soon, and model weights and system cards will be released in the coming days.

In February 2026, Jason Maynard succeeded Zig Serafin as CEO of experience management company Qualtrics. Nine months after taking office, he made a decision that seemed minor but was pointedly clear: banning the word 'headless' internally. "I won't allow anyone to put headless on any slide," Maynard told diginomica, "because that is not our product strategy, nor our vision." He coined an antithetical alias for the term: brainless.

This is not nostalgia for screens and dashboards. Maynard acknowledges that as AI agents take on more work, the scenario of people logging into applications to view reports is indeed diminishing. What he opposes is the architectural implication behind the word 'headless': it suggests you buy a system and can extract all its data and logic for external agents to consume at will. "If the head is gone, so is the brain," he said. In Maynard's view, this narrative surrenders control of customer relationships to third-party agents—personal AI assistants attempting to interpose themselves between brands and users—which brands are collectively blocking, precisely exposing the problem.

Qualtrics' alternative is called 'embedded insights'. Maynard anticipates that at some future point, 90% to 95% of the company's output will no longer exist in dashboard form, but rather as 'insight tokens' embedded where decisions happen—in CRM systems, customer service tickets, marketing automation platforms, and even third-party agents. In September 2026, Qualtrics released its new XM Data & AI Platform (scheduled for delivery in 2027), built around three capabilities: simulation (testing how a decision lands before committing to it), prediction (using small language models to anticipate specific individuals' likely reactions), and trusted outcomes (validating against organizational rules before any action reaches a real person). The key distinction is that data and reasoning logic remain inside the Qualtrics platform, sending out only the final, validated insight.

Maynard cited an airline example: a frequent flyer program deciding who gets an upgrade relies on a set of non-negotiable, deterministic business rules. The question is whether these rules should be distilled and embedded into the framework of a frontier large model lab? He believes most organizations do not want this, because "that is our intellectual property, important stuff, to keep within our own harness." This leads to what he calls 'cross-currents'—enterprise customers currently face two opposing forces: on one hand, the diffusion of experience expectations is accelerating faster than anyone anticipated; once consumers have a good experience somewhere, they assume all touchpoints should be like it. On the other hand, trust in AI is declining while demands for data sovereignty are rising. Both things are true simultaneously.

This is why Maynard insists on using 'trusted outcomes' rather than simply 'outcomes'. Qualtrics previously described what it does with three words: listen, understand, act. The new platform added two words: simulate, predict. But the qualifier 'trusted' was added later. "If I don't write trusted, just outcome, it makes no sense; if I just write trust, it also makes no sense," he said, "it must be those two things tied together." He added a remark with industry-critical undertones: "No organization—whether government or commercial—can stay in business supplying slop. Brainless, headless slop. Your business, done."

In the field of computer vision, although Convolutional Neural Networks (CNNs) have achieved great success, their computational costs are high. As pointed out in the paper "Recurrent Models of Visual Attention," applying CNNs to large images causes the computation to grow linearly with the number of pixels, resulting in extremely long training and testing times. To address this issue, researchers at Google DeepMind, inspired by human visual perception, proposed a new recurrent neural network model.

When humans observe a scene, they do not process the entire image at once; instead, they selectively focus attention on specific parts of the visual space, gradually accumulating information to build an internal representation. This mechanism not only saves "bandwidth" but also reduces task complexity by ignoring irrelevant backgrounds (i.e., "clutter"). The core contribution of this paper is modeling visual attention as a sequential decision-making process. The model, known as the Recurrent Attention Model (RAM), interacts with the environment like an agent. At each step, it extracts local high-resolution information through a bandwidth-limited sensor (Glimpse Sensor) and determines the next fixation point based on previous gaze positions.

Since the process of selecting fixation points is non-differentiable, the research team employed reinforcement learning for end-to-end training. Specifically, backpropagation was used to train the neural network components, while Policy Gradient methods were utilized to solve the non-differentiability issues in control problems. Experiments showed that this model significantly outperformed CNN baselines on cluttered image classification tasks and could dynamically track objects without explicit supervision signals. This method demonstrates the potential of attention-based models in handling complex backgrounds and scaling to large input images.

For researchers and engineers interested in understanding how to introduce reinforcement learning into visual systems or explore efficient image processing architectures, "Recurrent Models of Visual Attention" is an important classic literature. It clearly shows how simulating biological visual mechanisms can optimize the efficiency and robustness of deep learning models. It is recommended to start reading from the abstract and Section 3, "Recurrent Attention Model (RAM)," focusing on its architecture design and training process. Published in 2014, this paper remains a cornerstone for understanding the combination of attention mechanisms and reinforcement learning today.

In the 1995 letter to shareholders, Warren Buffett detailed Berkshire Hathaway's unique capital allocation philosophy. This letter not only recorded the year's financial performance but, more importantly, revealed how the company built its business empire through a "dual-track" strategy.

First, Buffett introduced a key concept: the double-barrelled approach. He stated that Charlie Munger and he aimed to build a collection of companies with excellent economic characteristics and outstanding management. Their preference was to acquire 100% ownership through negotiation at a fair price. However, they were equally satisfied when the stock market offered the opportunity to buy small stakes in outstanding businesses at prices below what would be required for full acquisition. This method of negotiating to buy entire businesses or buying partial interests via the stock market gave Berkshire an important advantage over capital allocators who stick to only one path. The core of this strategy lies in flexibility, allowing investors to capture the best value under different market conditions, whether through controlling stakes or minority investments.

Second, the letter emphasized the discipline of comparing acquisitions with passive investments. Buffett noted that when considering any action, they mentally compared it with dozens of other opportunities, including buying tiny shares of the world's best businesses through the stock market. This practice of contrasting acquisitions with passive investment is a discipline rarely used by managers focused solely on simple expansion. He cited Peter Drucker's view that transactions are often more attractive than work, but this can lead to meaningless deals. Therefore, one of Berkshire's advantages is having no fixed strategic plan, thereby avoiding paying foolish prices just to follow a predetermined direction. This strict comparison based on opportunity cost ensures that every capital expenditure undergoes the most rigorous value test.

For modern readers, understanding this mechanism helps reflect on the "anchoring effect" in personal portfolios or corporate decisions. Suppose an investor faces two choices: directly buying all shares of a startup and participating in operations, or buying stock in a listed industry leader. Without the discipline of weighing both options against each other in terms of opportunity cost, as Berkshire does, an investor might ignore the potentially higher risk-adjusted returns of the latter due to a pursuit of "control" or the "excitement of the deal." This mental framework reminds us that investment decisions should not be driven solely by emotion but by a rational assessment of alternatives.

The Berkshire Hathaway 1995 Shareholder Letter, written by Warren Buffett, was published on March 1, 1996 (reviewing fiscal year 1995). It is suitable for investors seeking to deeply understand long-term value investment strategies, capital allocation logic, and how to make rational choices between controlling and non-controlling stakes. It is recommended to start reading from the "Acquisitions" section of the letter, focusing on paragraphs regarding the dual-track strategy and opportunity cost comparisons.