Reading

When building enterprise-grade AI agent analytics, a recurring problem is that agents can write SQL but do not know the business meaning behind column names or what they are authorized to read. Common solutions involve deploying a separate semantic layer—such as a knowledge graph, data virtualization platform, or dimensional model—and having the agent query that layer. On November 3, 2026, the Amazon Web Services Machine Learning blog presented a different approach: placing the semantic layer directly inside the operational Postgres database itself, using native database COMMENT, GRANT, and row-level security features to define meaning and control access, eliminating the need for a second system.

The specific mechanism operates in three layers. The first layer is meaning: the AIDB extension’s Semantic Knowledge Base (SemKB) embeds the definition of every table, view, and column, along with COMMENT ON text, into a vector index, allowing agents to find the correct tables and columns via natural language search. The second layer is relationships: Postgres foreign keys are automatically crawled by SemKB into join paths, so agents do not need to guess associations based on column names. The third layer is governance: the AIDB Model Context Protocol (MCP) endpoint has no permission system of its own; it passes the Postgres username and password for each request as HTTP Basic authentication, and the database’s GRANT and Row-Level Security (RLS) policies determine what that connection can see.

The blog illustrates the role of the meaning layer with a concrete example. In the C360 dataset, the purchases.price column records the order total, not the unit price. Without a COMMENT ON statement clarifying this, the agent treats it as a unit price multiplied by quantity, calculating approximately $502,529 for the top customer, Carolyn Inworth—a wrong figure. Adding a comment stating 'Total amount the customer paid for the entire order, in US dollars. Not the unit price.' changes the answer to $27,991.66. The model did not change, the prompt was not altered, and the tools remained the same; only the indexed comment within SemKB changed.

Governance verification is more direct. The blog defines two Postgres login roles: mcp_agent_full can see all columns, while mcp_agent_restricted cannot see customers.ssn. Running the exact same prompt to list five customers' Social Security Numbers yields different results. The full role returns the SSN values for all five customers; the restricted role’s tool call is denied at the Postgres layer, returning 'permission denied for column ssn,' after which the agent informs the user that SSNs cannot be read. The prompt, tool catalog, SemKB, and agent code are identical; the only difference is the Basic authentication credentials injected by AgentCore Identity.

Architecturally, the agent runs on the Amazon Bedrock AgentCore runtime, using Claude for planning and calling MCP tools through the AgentCore Gateway. The Gateway forwards requests to an Application Load Balancer within the EDB Postgres AI Virtual Private Cloud (VPC), reaching the AIDB edb-endpoints MCP server. No Lambda functions or translation layers are required—AIDB natively speaks the streamable HTTP MCP protocol. Auditing also remains within the same database: AIDB writes OpenTelemetry traces for every SQL execution into the aidb_otel schema, which can be queried using standard SQL.

On November 3, 2026, OpenAI announced the launch of a program called "Grid Protectors," aimed at reducing the power consumption of its data centers during peak hours to ensure local grids are not overwhelmed. According to Tom's Hardware, this move is intended not only to safeguard residential electricity but also to allow its data centers to connect to the grid and go online sooner, without waiting for additional power sources to come online.

OpenAI is testing the system at its planned campus in Effingham County, Georgia. The company stated that it has already signed a 3.2-gigawatt agreement with Georgia Power, which includes 1,000 megawatts of flexible demand response. This means that on the hottest days of the year, if electricity demand surges, this 1-gigawatt buffer would be enough to power approximately 800,000 households, equivalent to nearly 20% of the state's homes.

This technology, known as demand response, has been around for decades, typically involving grid operators manually calling up large industrial customers like steel mills to cut consumption. Today, with modern technologies and IoT, data centers can automatically reduce their load in under a minute. Nvidia has previously experimented with a similar system in the UK and proven its effectiveness. OpenAI stated that it intends to make this kind of flexibility a core part of how it builds and operates future campuses, working with utilities to agree on how much demand can be reduced, how quickly, and for how long.

US tech giant Google announced on November 3, 2026, that it plans to default-disable AI tools for K-12 students next year unless school administrators explicitly turn them on. The change affects Gemini, NotebookLM, Gemini in Classroom, and Google Search.

The critical constraint is that because AI features are deeply integrated into the search engine, schools choosing to restrict AI altogether will no longer be able to access Google Search. Chris Phillips, Vice President for Education at Google, told The New York Times that "Search and AI are deeply integrated," meaning if a school chooses to restrict AI, students "will not have access to search, and they’ll see a message from their school that indicates it’s not available."

This creates a dilemma for schools. While teachers may not want students using AI Overviews in Google Search to conduct research or complete assignments, they likely still want them to have access to the search engine to find sources and other learning materials. However, Google Search and its AI features are now inseparable. Given the dominance of Google’s suite of apps in classrooms and the widespread use of Chromebooks, the decision underscores the company’s influence over the education sector. Schools that want to continue having access to Google Search will have to accept its AI features.

"We’ll share more specifics with schools in the coming months ahead of the changes rolling out so they have time to prepare and enable a smooth transition," Phillips wrote in a blog post. "Our goal is simple: provide school leaders the control and confidence to use AI products in classrooms as they see fit."

Google also launched Guided Learning alongside this announcement. Unveiled earlier this year, the tool functions like an AI tutor designed to help users build deep understanding rather than just getting answers, responding to concerns that AI chatbots undermine the learning process by providing direct answers.

Mirae Asset Securities announced on October 1, 2026, that it is officially renaming its mobile trading system (MTS), previously known as 'M-STOCK', to 'MAPS'. The platform has been upgraded to a comprehensive investment hub centered on portfolio management. This move aims to leverage artificial intelligence to optimize users' decision-making processes, shifting focus from traditional quote viewing and order execution to smarter asset allocation support.

The new MAPS platform highlights two core sections on its home screen: 'Discovery' and 'My Investment'. The 'Discovery' section introduces an AI service called 'Market Now', which automatically generates a one-sentence summary of market trends every two hours, helping users quickly capture key information. Additionally, the platform has added a digital assets menu providing relevant market data and professional research materials, while expanding coverage to include information on global innovative companies such as US non-listed stocks.

Mirae Asset Securities stated that the long-term goal of MAPS is to realize an 'AI Everywhere' environment, ultimately evolving into 'Agentic WM' (Agentic Wealth Management). In this model, AI will not only provide information aggregation but also proactively monitor market changes and suggest investment actions based on user goals and constraints. Currently, the company's internal Tech & AI division, comprising about 40 personnel, is advancing this strategy, with over 85% of this year's targets for internal efficiency and customer service reportedly achieved.

According to WIRED, independent research firm Forum AI found that nearly 40% of answers provided by mainstream AI chatbots—including OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini—contained at least one factual error when responding to questions about the US midterm elections. For instance, Gemini incorrectly stated the mail-in ballot deadline for Pennsylvania, and Claude confused North Carolina Congressman Don Davis with a former school superintendent of the same name.

Research by NewsGuard further indicated that these models frequently cite “pink slime” websites—low-quality sources funded by political or ideological groups but presenting as mainstream news—when asked about candidates and policies. ChatGPT cited such sites in more than 70% of cases, while xAI’s Grok performed better with a 30% rate.

Pew Research data shows that approximately 8% of US adults used AI tools to find information on candidates and policies ahead of the midterms. Despite OpenAI publishing election safety guidelines, the findings highlight significant reliability risks in current AI technology when handling real-time, complex political facts, potentially misleading voter decision-making.

Angle Health, a US health benefits technology company that administers health insurance plans for small and midsize employers, has an internal service team that fields thousands of calls daily from provider offices, approximately 80% of which involve eligibility verification, benefit inquiries, prior authorization follow-ups, and claim status checks. These repetitive questions previously consumed significant manpower. In November 2026, ElevenLabs published a customer case study detailing the pilot results of Angle Health deploying voice agents on its ElevenAgents platform to handle these calls: the agents took over about 80% of the total call volume, with more than 90% resolved end-to-end without transferring to a human agent, allowing the team to save nearly 9,000 minutes per day for complex escalation cases.

Abraham Duplaa, Head of Product at Angle Health, interprets these results differently from the typical automation narrative. He stated that what truly impressed him was not the 90% automated resolution rate, but the remaining 10%—the agents' ability to identify that they could not handle a specific call, understand why, and route it to the correct team. In his view, the key to whether a voice agent is viable in a medical setting is not how many calls it processes, but whether it can accurately judge which calls it should not handle and send them to the right people. This perspective shifts the deployment standard for voice agents from simple automation coverage to a focus on escalation quality: the system must not only be able to answer questions but also know what it cannot answer.

From a compliance standpoint, Angle Health's pilot operates solely within the US HIPAA framework. All voice agent deployments require a signed Business Associate Agreement (BAA) and utilize ElevenLabs' Zero Retention Mode to ensure patient data is not stored. Agents must accurately capture clinical terminology and diagnosis codes during conversations, which fundamentally distinguishes this use case from general customer service scenarios. Caleb Engelbourg, Head of R&D & Strategy at Angle Health, stated that the team compared multiple voice agent vendors during the pilot, and ElevenLabs demonstrated the best responsiveness accuracy in medical settings.

In production, Angle Health employs deterministic workflows to log necessary information and enforce handoff rules, while also using the built-in call review feature of ElevenAgents to confirm that these rules are being followed correctly. This means the agents' behavior is not entirely black-boxed—the team can trace back each call to verify if it ran according to preset rules. When an escalated claim is resolved, the agent proactively calls the provider office to inform them of the outcome, rather than waiting for the office to call again to inquire.

Angle Health currently plans to expand this pilot to more business segments. Duplaa stated that the team hopes to extend the partnership to commercial, brokerage, and account management departments, enabling all stakeholders interacting with Angle Health to receive answers faster.

Korean edtech company Weverse Brain officially launched its AI agent platform "AgentMax" at its "2027 Tech Day" event in Seoul on November 3. The platform is positioned as an "incubator" for developing specialized AI agents across various fields, distinguished by its focus on "behavioral AI" that actively guides users toward completing specific actions, such as sticking to a study plan or making a purchase decision, rather than just engaging in conversation.

The core technology behind AgentMax is the EDRA dialogue design engine, which stands for Empathy, Diagnosis, Reframe, and Action. Company CEO Jo Se-won stated that the engine leverages ten years of phone-based learning data and conversations from 1.23 million cumulative members to identify user states and stimulate their willingness to act. According to company-reported data, the learning retention rate three months after joining was 63% for users who signed up between September 2025 and June 2026, a figure described as rare in the education industry.

Live demonstrations featured two representative agents: the AI English tutor "MaxTutor" and the AI sales consultant "MaxSales." MaxTutor proactively calls busy students to suggest practicing just one sentence, while MaxSales addresses price concerns in a simulated sales scenario by converting monthly payments into daily costs and manipulating the browser interface in real-time to show competitor comparisons and promotional offers.

Weverse Brain plans to expand its business through a B2B2C model. Current partnerships include supplying AI tutors to NE Neungyul and signing an MOU with Hankuk University of Foreign Studies to develop multilingual AI tutors. The company emphasized that R&D funding comes from existing business profits, reporting last year's revenue at 41.5 billion KRW and operating profit at 3.2 billion KRW.

In 2021, the Google Research team published "Finetuned Language Models Are Zero-Shot Learners" (the FLAN paper), proposing a simple method to significantly improve the zero-shot learning capability of language models through "instruction fine-tuning." The core of this research lies in converting natural language processing tasks into formats with explicit instructions, such as "Is the sentiment of this sentence positive or negative?" or "Translate 'Hello' into Chinese," and fine-tuning the model on more than 60 such datasets.

Traditional large-scale pre-trained models (such as GPT-3) perform well under few-shot prompting but often struggle to understand tasks that differ significantly from pre-training data formats in zero-shot scenarios. The FLAN paper points out that by teaching models to follow instructions through supervised fine-tuning, they can generalize to unseen task types. Researchers grouped NLP datasets by task type; during training, one group of tasks was excluded, and the model's zero-shot performance on that group was evaluated only at the testing stage. Experimental results showed that the 137B-parameter FLAN model, after instruction fine-tuning, surpassed the 175B-parameter GPT-3 zero-shot performance on 20 out of 25 unseen tasks, and significantly outperformed GPT-3's few-shot performance on some reasoning and reading comprehension tasks.

This finding reveals three key factors important for the success of instruction fine-tuning: the quantity of the fine-tuning dataset, the model size, and the natural language instructions themselves. For current AI developers, FLAN demonstrates a path to stimulate the latent capabilities of models through structured instructions without requiring massive amounts of labeled data. Readers are advised to start with the abstract and introduction of the paper, focusing on the methodology comparison shown in Figures 1 and 2, to understand how instruction fine-tuning combines the advantages of the pre-training-fine-tuning paradigm and the prompting paradigm. This paper, published in 2021, titled "Finetuned Language Models Are Zero-Shot Learners," laid an important foundation for the instruction-following capabilities of subsequent large models.

In the 2023 shareholder letter, Warren Buffett first mourned the recently deceased Charlie Munger and reviewed Munger's influence on Berkshire's investment philosophy. Buffett recalled that in 1965, Munger bluntly advised him: "Warren, forget about ever buying another company like Berkshire. But now that you control Berkshire, add to it wonderful businesses purchased at fair prices and give up buying fair businesses at wonderful prices." This advice prompted Buffett to adjust his strategy, but the original text explicitly noted that this process was not instantaneous; rather, it involved backsliding and struggle: "With much back-sliding I subsequently followed his instructions." Munger later became Buffett's partner and, when Buffett's old habits resurfaced, he "repeatedly, jerked me back to sanity." This shift from Graham-style "cigar-butt" investing to holding high-quality assets was described by Buffett as a gradual process requiring constant vigilance, rather than a simple break. Consequently, Munger was called the "architect" of Berkshire by Buffett, while Buffett viewed himself as the "general contractor" executing that vision.

Beyond the historical review of strategic transformation, the letter also clarified financial concepts often confused by investors, particularly the difference between GAAP net earnings and operating earnings. Buffett pointed out that "Net earnings (loss)" in official annual reports includes unrealized capital gains or losses, which fluctuate wildly and often mislead readers. For example, GAAP net earnings for 2021, 2022, and 2023 were $90 billion, negative $23 billion, and $96 billion respectively, with such dramatic volatility primarily stemming from daily stock market fluctuations. In contrast, Berkshire places greater emphasis on "operating earnings," which represent core business profitability excluding unrealized capital gains and losses. The corresponding figures for the same period were $27.6 billion, $30.9 billion, and $37.4 billion, showing a steady growth trend. Buffett emphasized that while capital gains are an important component of Berkshire's value appreciation, judging a company's investment value based on "earnings" that include short-term market volatility is absurd. He quoted Benjamin Graham's famous saying: "In the short run the market acts as a voting machine; in the long run it becomes a weighing machine," reminding investors to focus on the accumulation of intrinsic corporate value rather than instantaneous stock price fluctuations.

For ordinary investors today, this letter provides a clear reading framework: how to see through complex accounting noise to the essence of business. Suppose an investor faces a tech giant whose financial statements show significant net income volatility due to changes in the fair value of stock options. If they only look at GAAP net earnings, they might make incorrect judgments. Borrowing from Buffett's approach, the investor should strip away non-operating factors and focus on cash flows and operating earnings generated by the company's main business to assess whether it possesses the qualities of "wonderful businesses." This analytical method applies not only to Berkshire but also to any high-quality enterprise one hopes to hold long-term. It requires investors to ignore short-term market sentiment (the voting machine) and instead focus on long-term fundamental performance (the weighing machine).