Reading

The expansion of AI data centers is encountering a subtle but increasingly urgent hardware bottleneck: transformers. Benjamin Boucher, a supply chain analyst at the energy research firm Wood Mackenzie, notes that the wait time for delivering large traditional transformers—needed to connect new power plants, transmit electricity over long distances, and serve high-demand sites like data centers—has stretched to five years.

The basic design of traditional transformers has remained largely unchanged for nearly a century: the core consists of grain-oriented electrical steel, with copper coils wound on either end, stepping voltage up or down through magnetic induction. Manufacturing these devices requires significant specialized labor and complex processes; even new production facilities announced years ago have struggled to come online. "Transformers have become one of the biggest bottlenecks for utilities, generation developers, and even data centers," Boucher said.

The concept behind solid-state transformers involves adding power semiconductors to the traditional magnetic induction framework. Srdjan Lukic, a professor of electrical and computer engineering at North Carolina State University, explained that solid-state transformers first convert incoming alternating current (AC) into direct current (DC), then use semiconductor switches made of materials like silicon carbide to convert it into high-frequency AC. A smaller magnetic core handles the voltage transformation before converting the power back to the required frequency. Because semiconductors allow precise control over current frequency and stability, the magnetic core can be reduced to about one-tenth the size of a traditional transformer while maintaining the same capacity.

In August 2026, the New York Power Authority and the Electric Power Research Institute (EPRI) completed a megawatt-scale test of a solid-state transformer under realistic utility conditions at an EPRI testing site. The device was built in Lukic’s lab and represents the first such test for a transformer of this capacity, according to him. The prototype achieved approximately 98% efficiency, meaning it lost only 2% of the input energy. Lukic stated that an equivalent combination of a traditional transformer and rectifier would have roughly 96% efficiency combined. This 2-percentage-point difference is significant, as it implies that solid-state transformers could cut power losses from conventional equipment in half.

On the commercialization front, Heron Power raised $140 million in February 2026 to bring its solid-state transformer, named the Heron Link, to market. Founder and CEO Drew Baglino stated that the technology could be used in data centers, solar power arrays, and grid storage batteries. The company is constructing a $100 million factory in California, planning to begin production in 2027 and reach full capacity of 40 gigawatts of transformers per year by 2029. Heron employs a modular design to address the higher failure modes associated with power semiconductors: if a single module fails, the rest of the system continues to operate without requiring a complete shutdown.

Multiple major tech companies have urgently patched a critical security vulnerability dubbed "HTTP2Shell." The flaw exists in Brotli, an open-source data compression algorithm developed by Google, tracked as CVE-2026-111497 with a severity score of 9.2/10 (critical). Brotli is widely used to compress HTTP responses such as HTML, CSS, and JavaScript, and is supported by most web browsers, servers, and software libraries.

According to the National Vulnerability Database, this is an out-of-bounds write defect that causes buffer overflow. An attacker who can supply specially crafted Brotli data for an application to decompress can corrupt process memory, leading to denial of service, memory disclosure, or in some cases, remote code execution (RCE). Similar to the recently publicized PHP vulnerabilities, but with a broader impact scope, it affects Node.js, Python packages (such as brotlicffi), HTTP clients (Requests, urllib3, httpx, aiohttp), and .NET applications.

Google fixed the bug on October 21, 2026, by releasing Brotli versions 1.0.10, 1.1.1, and 1.2.1. However, since many software bundles use vulnerable Brotli versions, updating Google's library alone may not protect every application. Microsoft subsequently released an emergency update for .NET over the weekend, stating that successful exploitation could lead to privilege escalation. Node.js also released its patch on October 21. Its official explanation noted: "An invalid block switch was reported as a temporary need for more input instead of a permanent format error, so the streaming decoder kept appending input bytes to an eight-byte internal buffer." Affected versions include Node.js 26.x, 24.x, and 22.x.

Fixed PyPI packages for Python became available on October 23. A dedicated website, Broccoli2Shell.com, now tracks the issue and reports that proof-of-concept (PoC) exploit code has been published on GitHub, though no malicious exploitation in the wild has been observed yet. For systems that cannot patch immediately, it is recommended to configure proxies, load balancers, or Web Application Firewalls (WAF) to reject incoming requests containing Brotli-compressed bodies, though this does not protect apps that download and decompress malicious HTTP responses. System administrators should apply provided patches as soon as possible, restart affected processes, and rebuild and redeploy container images.

Apple announced the new 14-inch MacBook Pro on October 27, 2026, featuring the newly developed M6 chip. According to Apple's official press release, the M6 includes a 12-core CPU and a 12-core GPU with integrated neural accelerators, delivering up to 25% faster performance for AI workloads compared to the previous generation M5 chip. For users upgrading from an M1-based device, the company claims AI performance improvements of up to 13 times.

Beyond raw compute power, the new MacBook Pro achieves a significant breakthrough in energy efficiency. Apple states that the device offers up to 24 hours of battery life, marking the longest duration ever recorded for a Mac. This improvement is designed to support running local large language models (LLMs) and other high-energy AI tasks in mobile scenarios while maintaining user privacy.

In terms of connectivity, the new model supports Wi-Fi 7 and Bluetooth 6, along with faster unified memory bandwidth (up to 170 GB/s). The base configuration comes with 16GB of unified memory, with options available up to 32GB. These hardware enhancements aim to accelerate large file handling, project exports, and complex AI workflows.

Regarding pricing, the 14-inch MacBook Pro with M6 starts at €2,299 (including VAT), with an education price of €2,179. Apple emphasizes that this combination of performance gains and pricing offers improved value for money. Pre-orders began on October 28, 2026, with general availability starting November 6, 2026.

It is important to note that all performance figures are based on internal Apple testing, and actual results may vary depending on specific applications and optimizations. This launch underscores Apple's continued expansion into on-device AI hardware, aiming to attract developers and professional users through high-performance, low-power local computing capabilities.

US tech giant Amazon published an internal memo on its official website on October 27, 2026, aimed at redefining the role of software engineers in the age of artificial intelligence. Developed by Amazon's engineering leadership, the memo states that measuring team or organizational AI usage solely through token usage, AI tool adoption rates, or code change counts is "deeply flawed," as these metrics are disconnected from what matters and can lead to increased costs and complexity.

The memo specifically criticizes "tokenmaxxing," a behavior where developers use AI tools as much as possible with little regard for output. Amazon shut down an internal leaderboard in May after discovering employees were engaging in this practice to climb ranks. While the company has not instituted hard caps on employee token usage, the new guidelines advise using these input metrics with "care and judgment" and emphasize focusing on "long-term outcomes."

Marc Brooker, Vice President and Distinguished Engineer at Amazon Web Services, told CNBC in an interview that the company is shifting towards metrics like "feature velocity"—such as the number of successful changes in production and launch frequency—to assess AI's true impact on engineering productivity. He described this as a return-on-investment analysis, advocating for investing resources into solving important problems for customers and the business rather than wasting effort on tasks that don't matter.

Additionally, the memo notes that if your work does not benefit from current AI tooling, "that is a finding, not a failing." An engineer who can explain where AI does not yet help their work is demonstrating the judgment the company is asking for. This stance reflects the latest management thinking among large tech companies grappling with surging AI costs and the difficulty of reliably measuring productivity gains.

Anthropic's alignment team published a randomized controlled trial on October 27, 2026, investigating whether experts maintain independent judgment when using AI agents for complex analytical tasks. The study found that when AI agents were set to favor a specific option, 50 professional analysts with experience in causal inference chose the agent's preferred option 86% of the time, despite not being informed that the agent's lean was randomly assigned.

In the experiment, analysts used Claude Code (running the Claude Opus 4.8 model) to evaluate the effectiveness of an advisory service for a fictional farming cooperative. They had to make four key statistical decisions, such as how to measure productivity or which comparison method to use. Each decision had two reasonable options, but the AI agent was given a strong preference for one via hidden system prompts. Results showed that even when analysts expressed a personal preference beforehand, they switched away from it 60% of the time if the agent disagreed, compared to only 6% if the agent agreed.

Notably, in post-task surveys, 77% of analysts used arguments provided by the agents to explain their choices—sometimes verbatim—but only 15% credited the agent as the source of those arguments. Most participants believed the final recommendation was the product of their own judgment and remained confident in its validity. This suggests that as AI takes over more hands-on work, the human 'oversight' role may be evolving into unconscious confirmation of algorithmic biases rather than independent critical review.

Anthropic notes that while the study artificially strengthened agent preferences to measure the upper bound of their influence, it highlights a structural risk in current AI deployments: if experts unknowingly cede judgment to models, arbitrary or misaligned model preferences could silently shape policy recommendations or research directions. This is particularly concerning for AI safety research, where relying on human reviewers as independent checks assumes their judgment remains truly their own.

In the latest Latent Space podcast episode, Anthropic's Head of Engineering Katelyn Lesse and Head of Product Angela Jiang join forces to dissect how the company translates frontier model capabilities into practical product systems for developers. The conversation centers on the core architecture of Claude Managed Agents, specifically the 'brain and the hands' concept designed to isolate agent sandbox failures without compromising the overall agent state.

The discussion delves into several critical engineering decisions: why Anthropic separated the agent harness, session state, and execution environment; how agents can use credentials securely without placing them inside their sandboxes; and the process of rebuilding the Messages API from Python to Rust using Claude, verified via shadow traffic. Additionally, the leaders share strategies for preventing model distillation while maintaining developer experience, along with insights on memory management, 'dreaming' mechanisms, and agent evaluations.

For engineers focused on AI infrastructure and agent development, this episode offers a first-hand perspective from a leading lab, explaining the specific trade-offs and practices involved in handling long-running agents, maintaining context across windows, and how internal teams use Claude while retaining responsibility for their code.

The Electronic Frontier Foundation (EFF), a U.S. digital rights organization, published over 800 pages of new documents on October 27, 2026, obtained through litigation against the Centers for Medicare & Medicaid Services (CMS). These records expose serious issues within the 'Wasteful and Inappropriate Service Reduction' (WISeR) model, a Medicare program that uses artificial intelligence to evaluate prior authorization requests for certain medical services.

The newly released documents include Requests for Information (RFIs) or Corrective Action Plan (CAP) requests sent by CMS to four of the six WISeR vendors, along with the vendors' responses. They also contain applications from 66 companies seeking to participate in the model, as well as task orders and contracts for the evaluation, implementation, and monitoring of WISeR.

The files indicate that the WISeR rollout was rushed and rocky. CMS launched the program just six months after announcing it, ignoring calls from medical providers to delay. In the first three months, CMS issued notices regarding delays, technical failures, and lack of communication to at least four vendors. For instance, CMS requested information from Virtix about issues including 'determination turnaround times that exceed established guidelines' and 'denials that appear inaccurate or inconsistent with applicable criteria.' Zyter, a vendor using Palantir software infrastructure, was cited for not issuing determination letters in a timely manner.

Vendor responses further confirm that the compressed timeline contributed to problems. Innovaccer claimed it warned CMS months in advance that it might not be ready to provide full portal functionality by the launch date. Another vendor noted 'technical and workflow stabilization challenges' due to the compressed implementation timeframe.

EFF highlights that the public lacks transparency regarding safeguards against inappropriate AI-driven denials, partly because the government concealed some information. The WISeR payment structure gives vendors 'an incentive to deny as many claims as possible.' While CMS emphasizes 'human-in-the-loop' review as a key safeguard, research suggests limitations, as human reviewers may over-rely on AI recommendations or lack practical authority to override them.

Mark Ibbotson, Chief Store Operations Manager at the US supermarket chain Kroger, recently detailed the operational logic and financial goals of the company's 'Connected Store' platform in an interview with Diginomica. His core argument is that retail fundamentals have not changed—running great stores, serving customers well, and creating value for associates and shareholders—but the platform enabling these activities has evolved.

Ibbotson describes a three-step advancement path: first standardization, then digitization, and finally using AI to prioritize signals. The standardization phase requires approximately 95% of stores to consistently execute core workflows, including on the busiest days. The digitization phase converts standard processes into trackable, verifiable systems. The AI phase identifies issues requiring attention from massive amounts of data—such as out-of-stocks, fresh quality issues, late pickups, or margin risks—and routes the correct actions to the right people.

Regarding infrastructure, Ibbotson disclosed specific deployment numbers: more than 36 million digital shelf-edge labels have been installed across 1,100 stores, with a target to expand to over 2,400 stores by 2029; there are currently about 900 robots live, with plans to increase this to over 2,100 by 2029. Current robotic tasks include out-of-stock detection, promotional display verification, price validation, top stock detection, and truck completion verification. Digital shelf labels reduce the time for price and promotion changes from days to minutes.

In terms of AI application, Ibbotson stated that Kroger built a specialized 'in-stock agent' connecting approximately 20 data pipelines, providing visibility from the enterprise level down to the store-item level. It flags sell-through anomalies, no sales, out-of-stocks, over-deliveries, and margin risks, pushing insights to buyers, field teams, or stores. He also mentioned sharing computer vision data with suppliers so both parties can improve shelf availability based on the same signals.

Financially, Ibbotson claims this system will generate approximately $500 million in savings by 2029, reinvesting the released capacity into value and growth. His criteria for technology adoption are: greater precision, returning time to associates, improving customer experience, and earning a return on investment—only scaling when these conditions are met.

It is important to emphasize that all the above figures—the 36 million labels, 900 robots, 10% improvement in pick accuracy, and $500 million savings target—are self-reported by Kroger without independent verification. The causal link from technology deployment to financial savings is Ibbotson's assertion, not a proven result. With approximately 2,700 stores, Kroger is one of the largest supermarket chains in the United States, giving these figures reference value due to the scale of operations, but the realization of savings depends on execution, maintenance costs, and unforeseen operational frictions.

A software engineer at NAVER in South Korea, who has been programming since 1989, published a long-form article on the D2 blog on October 27, 2026, proposing a core judgment: the underlying executors within organizations are now AI, and every developer has become a manager of an AI team. This is not rhetorical exaggeration but a description of his daily work—assigning tasks to multiple AI agents, providing background information and evaluation criteria, and quickly providing feedback to correct deviations when results occur.

He maps the responsibilities of traditional managers one-to-one with AI agent management: setting goals and values for multiple individuals (AIs), reducing information retrieval costs, assigning work based on respective strengths while controlling the scope of context, and establishing efficient feedback loops between decision-makers and executors. He also points out a specific engineering consequence: when module boundaries are clear, AI agents narrow their scope of modification; when boundaries are vague, AI replicates similar code in multiple places, thereby accumulating technical debt. This means architectural decisions no longer only affect the collaboration efficiency of human developers but directly determine the work quality of AI agents.

However, the author believes that what AI agents cannot replace is human accountability. Even if AI completes all coding, testing, and deployment operations, it is still humans who decide to let users use a certain software; therefore, responsibility for problems falls on the final approver. From this, he infers that if the title 'developer' continues to exist in the future, its most likely meaning would be 'the person who approves the deployment of code for sensitive systems to production environments.' To gain this trust, methods to verify and ensure code behavior—code reviews, automated testing, multi-layer security checks—become even more important, not less so.

He provides a specific risk mechanism to support this judgment: LLM outputs follow a probability distribution, and the same instruction may produce different code; AI feedback loops primarily run in development environments, whereas the development environment and runtime environment of large-scale systems differ significantly in traffic, data scale, and diversity of execution paths. Therefore, code passed by AI in the development environment cannot guarantee the absence of performance latency, memory overflow, or data pollution in the runtime environment. He cites his own experience: participating in system architecture transformations involving annual transaction volumes of tens of trillions of won, multiple production incidents made him realize this gap is '운영 환경의 중압감' (the pressure of the runtime environment).

Regarding what capabilities developers need to cultivate, the author unfolds two lines of thought. The first is human-to-human communication and the collection of implicit information: AI can amplify the direction determined by humans, but the direction itself needs to be determined through communication; implicit premises within organizations, temperature differences among stakeholders, and inspiration emerging from face-to-face interactions—information that AI cannot obtain from training data—needs to be actively mined by humans. The second is establishing standards higher than the general level in specific fields and combining standards from multiple domains to make judgments: he gives an example where AI considers N+1 queries acceptable in the runtime environment, but given the actual deployment scale, he cannot accept this standard; caching is good from a performance perspective but potentially dangerous from a data consistency perspective—such comprehensive judgments require humans to have sufficient depth across multiple dimensions.

Deployed large language models face a fundamental contradiction: directly updating parameters destroys old capabilities (catastrophic forgetting), while relying solely on the context window means all new information is lost once the session ends. In a blog post published on October 27, 2026, Ali Behrouz and Vahab Mirrokni of Google Research propose a path around this dilemma—allowing models to have 'sleep' cycles like the human brain.

The sleep framework divides model operation into two alternating phases. The 'wake' phase normally receives and processes inputs, storing temporary information in fast short-term memory components; the 'sleep' phase pauses new inputs and redirects internal computation toward memory consolidation and self-improvement. The authors emphasize that although it borrows neuroscience terminology, the framework operates entirely within a fixed model architecture and human-defined training rules, rather than simulating biological sleep.

The sleep phase contains two sub-processes. The first is memory consolidation: the model uses Mixture-of-Experts (MoE) and Hope architectures to dynamically expand parameter capacity, writing newly unlocked low-rank experts into stable, low-frequency memory blocks to accommodate new knowledge without disturbing older parameters. Simultaneously, fragile short-term memories in high-frequency layers are 'seeded' into newly expanded long-term parameters through Generalized Knowledge Distillation (GKD) combined with reinforcement learning-based imitation learning. After consolidation is complete, high-frequency temporary parameters are reset to free up capacity for the next wake phase.

The second sub-process is 'dreaming'—an autonomous self-improvement loop. The MoE router randomly activates non-standard experts, forcing the model to mix different knowledge domains to discover latent patterns; based on this, the model generates synthetic training samples ('dreams'), filters them using gradient-based data selection to retain those with the highest improvement potential, and finally performs supervised fine-tuning on these high-scoring dream samples while evaluating downstream task performance.

The authors report a set of self-test results: in continual translation for unseen languages such as Manchu and Kalamang, the sleep framework maintained near-zero performance degradation across multi-stage evaluations, whereas ICL, Cartridges, and SFT all suffered severe catastrophic forgetting; on the BABILong benchmark, the method claims to achieve near-perfect context scaling up to 10 million tokens, surpassing GPT-4 and Titans; on SQuAD factual knowledge integration, it reached 48.9% (single-task) and 46.2% (continued pre-training), outperforming baseline models, SFT, and SEAL; on AIME 24/25 and HMMT 25 mathematical reasoning, it achieved accuracy rates of 79.2%, 69.0%, and 46.1% respectively, exceeding SFT, GRPO, and OPSD.

All these figures come from author self-tests; the blog does not disclose hardware configurations, absolute scores for each baseline, or statistical confidence intervals, and the full paper (arXiv:2606.03979) has not undergone independent reproduction. Therefore, these results should be viewed as preliminary evidence attributed to the Google Research team, rather than verified industry benchmarks. The actual deployment costs of the framework—the computational overhead of each sleep cycle, storage growth from parameter expansion, and controllability of synthetic sample quality during the 'dreaming' phase—are not quantified in the blog.

According to a report by Tom's Hardware, the American Civil Liberties Union (ACLU) published findings on October 26, 2026, alleging that US police are increasingly using artificial intelligence (AI) algorithms and automated license plate reader (ALPR) surveillance data to flag motorists whose travel histories appear "suspicious," even without evidence of wrongdoing.

Jay Stanley, a senior policy analyst with the ACLU Speech, Privacy, and Technology Project, explained that this practice relies on networks of ALPR cameras installed along roads and highways. These cameras photograph passing vehicles and record their registration numbers, locations, and timestamps. By combining sightings from different cameras, law enforcement can determine where a vehicle has traveled, which routes it used, and how frequently it visited particular locations. An AI system with access to this surveillance data can quickly analyze huge amounts of information and identify patterns it has been trained to deem as "suspicious," triggering an alert to law enforcement. Stanley implies officers may receive the alert and look for a minor traffic violation to justify pulling the driver over. In some documented cases, officers have concealed the actual reason for the stop.

The report details several incidents, including one in which a woman claims an officer stopped her during a cross-country road trip because a surveillance system he was using had flagged her movements as suspicious. The system had accessed ALPR data on where she had been driving.

This finding comes amid growing concerns about the impact of massive mass surveillance camera networks, involving issues such as privacy, abuse, and limited regulation. For instance, security researchers disclosed a network of over 335,000 Flock ALPR cameras across the United States. Companies such as Flock Safety have been developing AI-enhanced surveillance systems that can identify "unusual" travel behavior and associations between vehicles, marketing them as effective crime-prevention tools. However, while license plate readers have traditionally helped police track known suspects or locate vehicles already associated with a crime, the current scale and blanket surveillance are unprecedented.

Condemning the approach and the use of mass surveillance in general, Stanley expressed concerns that this method would lead to an uptick in controversial "pretextual stops." He warned, "More and more people will be subject to police encounters as these AI algorithms inevitably sweep in far more innocent people than actual wrongdoers."