TAG40@@

TAG42@@TAG44

As usual, here is a curated selection of technical news and articles I found online today that personally caught my attention. I’d be happy to hear your opinions or any additional information in the comments.

Programming

.NET 11

Microsoft has announced the release of “.NET 11.” This release comes with two years of support as Standard Support (STS). It aims to make application development faster and more productive, with updates across a wide range of areas, including languages, the platform, and AI integration.

Advances in Languages and Tooling
In C# 15, the new default language, features such as the “Union type” and “closed hierarchy” have been added to make it easier to represent application state, while the compiler’s exhaustiveness checking has been enhanced. The Native AOT CLI host and MSBuild server are also enabled by default, reducing overhead such as startup times in development workflows.

Runtime and Performance
Numerous optimizations have been made, and improvements to the JIT compiler have enhanced the performance of existing code. In addition, “Runtime Async,” a feature that clarifies the call stack of asynchronous operations and makes debugging easier, has been introduced as a preview.

Strengthened AI Integration
For developing applications with AI built in, Microsoft provides capabilities such as “MEAI” for working with models, MCP for integrating with external tools, and MAF for building agents and workflows. Agent UI components for creating advanced AI experiences with Blazor have also been added on an experimental basis.

Libraries and Data Access
Standard support for Zstandard compression, enhancements to the Process API, and the LINQ FullJoin method have been added. In Entity Framework Core 11, SQL translation for queries has been optimized to reduce unnecessary JOINs, and experimental support for vector search with SQL Server has also been added.

Web and Client UI
ASP.NET Core now uses OpenAPI 3.2 by default and supports asynchronous validation..NET MAUI adopts CoreCLR as the default runtime across all platforms, while Windows frameworks such as WinForms and WPF also include support for the latest visuals and reliability-enhancing improvements.

Because .NET 8 and .NET 9 will soon reach end of support, a planned migration to the long-term support (LTS) version .NET 10 or the latest version, .NET 11, is recommended based on project requirements.

Impressions

Yay!

Rider 2026.3 has been released, introducing support for .NET 11 and C# 15, along with JetBrains Air, an autonomous development environment.

With JetBrains Air, you can instruct agents and validate their work from a new welcome screen without opening a solution. AI capabilities have also been significantly enhanced, including new skills for agents, test execution through Rider MCP, and copying project settings into Git worktrees to streamline parallel work and code reviews.

On the language front, in addition to support for shared types (unions), closed hierarchies, and extended indexers in C# 15, Rider now supports debugging .NET’s new Runtime Async.

Game development has also seen improvements in many areas, including professional debugging features for Godot, scene preview display, the addition of Lua and Verse plugins, and enhancements to the C++ debugger.

.NET Conf 2026

The .NET Conf 2026 keynote looked back on the platform’s 25-year journey while announcing the general availability (GA) of .NET 11 for the AI era.

Key updates include the introduction of Union Types in C# 15, faster processing through runtime async, and memory safety rules v2 to improve security. In addition, .NET MAUI was unified on CoreCLR, and the C# Dev Kit was made lighter and faster in preparation for the concurrent operation of multiple AI agents.

.NET Framework

This article summarizes the service updates for .NET and .NET Framework in November 2026. It includes important security fixes and release notes for multiple versions.

The Go Programming Language

The Go programming language marked 17 years since its open-source release. Over the past year, Go 1.26 and 1.27 were released, adding support for generic methods, simplifying syntax, introducing the new `encoding/json/v2` package, and adding goroutine leak detection.

Further hardware utilization and security enhancements are underway, including performance improvements through the new GC, the introduction of SIMD packages, and support for post-quantum cryptography. Going forward, the Go team plans to expand tooling, improve language features, and strengthen its community-driven approach.

Go versions 1.27.4 and 1.26.11 have been released, including 12 security fixes addressing issues such as HTTP/1 server crashes and improper code execution related to cgo.

As a special measure to address specific security vulnerabilities in Go 1.25, which is no longer supported, a Microsoft build of Go 1.25.14-2 has been released.

In a server written in Go, if a termination signal (SIGTERM) is received during a long-running operation, the operation may be forcibly terminated, potentially resulting in data inconsistency. To prevent this, use “Graceful Shutdown,” which stops accepting new requests and waits for in-flight operations to complete.

However, simply calling the Shutdown method does not communicate the shutdown status to request handlers. This is because, by default, the context held by a handler is not linked to the termination signal received by the main function. To address this, the main context must be passed to the server configuration’s BaseContext.

This enables handlers to detect context cancellation and safely exit during processing. For post-processing that must reliably be performed even in the event of an interruption, such as recording data to a database, context.WithoutCancel can be used to decouple the cancellation state and safely complete the shutdown process.

Developer Trends and Characteristics
Go developers include many highly experienced engineers, with 66% having more than six years of development experience, compared with an overall average of 49%. Annual salaries also remain high, and 70% use Go not only for business purposes but also for personal and side projects. Furthermore, rather than limiting themselves to a single language, 43% use TypeScript alongside Go in the same project, indicating that the combination of a Go backend and a TypeScript frontend has become established. While 70% of developers are not considering migrating to another language within the next year, Rust continues to be highly popular as a language to learn next.

Changes in Frameworks and Development Environments
In web development, the use of the standard library’s net/http surged from 32% in the previous year to 46%, approaching long-time leader Gin (48%). In addition to enhancements to routing functionality since Go 1.22, the trend toward choosing the standard library, which has fewer external dependencies, has grown as AI takes on the generation of routine code. In IDE environments, GoLand (49%) and VS Code (48%) are neck and neck, while the emerging Zed is gaining market share.

The Adoption of AI Tools and the Role of Humans
Go engineers are extremely proactive in using AI coding agents, with 86% using them via a CLI or terminal at least once a week. Fifty-five percent of developers said that more than 60% of the code they write for work was generated by AI. At the same time, rather than accepting AI output unquestioningly, 79% always have it reviewed by a human before adoption, carefully scrutinizing its functional correctness, quality, and design.

The Evolution of Go 1.26 and 1.27
The language itself has also evolved significantly. This includes the introduction of generic methods for concrete types, errors.AsType for type-safe error unwrapping, the implementation of UUIDs in the standard library, and the official integration of encoding/json/v2 into the standard library.Furthermore, simply migrating significantly improves runtime performance by making the new “Green Tea” garbage collector the default and optimizing small-object allocators.

C++

ReSharper C++ 2026.3 has been released.

As its headline feature, a new “Class Layout” window has been added to visualize the memory layout and padding of classes.

Additional improvements include enhanced format-string parsing for C++20/23 and the {fmt} library, refactoring enhancements such as changing constructor signatures, new CUDA-related inspections, and support for the latest C++23/26 features.

Rust

This article explains how to create a desktop To-Do application using GPUI, a high-performance UI framework written in Rust. GPUI was developed for the code editor Zed and is known for its high performance through GPU rendering.

The article provides a step-by-step guide to practical implementation using GPUI Kit. It comprehensively covers topics such as displaying windows, rendering the UI, managing application-specific state with entities, redrawing the screen through notifications, and configuring keybindings.

Git

Updated versions of Git have been released to address six security vulnerabilities, including arbitrary code execution during recursive cloning and memory corruption. Their impact spans a wide range of operations, including merging and applying binary patches.

Updating to the fixed Git 2.56.1 release or the latest maintained versions is recommended. If you cannot update immediately, take precautions such as avoiding recursive cloning and merge operations on untrusted repositories.

Visual Studio

Microsoft has announced the “LTSC 18.11” long-term servicing channel for Visual Studio 2026. In response to user requests, the support period has been extended from the previous one year to two years, through November 14, 2028.

This release introduces a new support channel for the existing product and includes LTS toolsets such as .NET 10. Although the Stable channel is recommended for access to the latest features, LTSC is the ideal choice for organizations that need to standardize their development environments.

Copilot Chat has received a major upgrade to a new engine in the latest version of Visual Studio 2026. By adopting the new Copilot SDK, it can now solve a wider range of coding tasks while continuing to use the same AI models.

New features also include shorter and more precise responses, the ability to change instructions while a task is in progress, detailed inspection of execution steps, and parallel work by subagents. These improvements further increase developers’ productivity.

MySQL

MySQL versioning transitioned to calendar versioning in the YY.M.P format, with 9.7 being the final release under the previous system. As a result, it is no longer possible to identify LTS releases from the version number, and a new “compatibility lineage” rule now applies to upgrade paths.

As a new feature, a new replication apply engine was added in 26.7. The subsequent 26.10 release introduced four-part version identifiers and componentization of REST Service, among other changes.

Android

Google has released a new open-source authentication plugin for Unity.

It directly supports Android Credential Manager and iOS GIDSignIn, enabling fast and secure sign-in even in Unity 6.5 environments. After obtaining a token, it can be integrated flexibly with various backends.

Claude Code

This article introduces five useful tips the author currently uses to leverage Claude Code more effectively, including commands, dedicated mods, and multi-agent configurations.

This article explains how to preserve the quality and safety of Claude Code outputs by recording rules, handoff information, and decision logs in files, and protecting sensitive data with a masking tool.

GitHub / GitHub Copilot

Google’s latest model, “Gemini 4 Carbon,” is now generally available in GitHub Copilot. It is designed for complex coding and reasoning, as well as secure development.

It is available to users of the Pro+, Max, Business, and Enterprise plans, and will become available in various development environments, including VS Code, in stages.

Anthropic’s latest model, “Claude Fable 5.2,” is now generally available in GitHub Copilot for complex coding tasks involving large codebases.

Starting December 8, 2026, GitHub will stop issuing Dependabot alerts for repositories that have been inactive for the past year and archive existing alerts.

MSBuild compile-command extraction has been integrated into the Microsoft C++ Language Server for GitHub Copilot CLI, enabling the compilation database to be generated automatically when a project begins without requiring any prior configuration skills.

Papers / Other

Software engineer Jameson Hodge has released “CHIMERA Vocal Synth,” which synthesizes speech by physically modeling the vocal organs.

Its distinguishing feature is that it creates sound without using recorded data, calculating the movement of air through the lungs, vocal cords, and airways. In addition to spoken and sung English, it can also reproduce animal sounds.

The audio processing engine is implemented in Rust and runs in the browser. It is currently publicly available as open source.

AI

Google

An article introducing the general availability of Google’s latest frontier model, “Gemini 4 Carbon,” along with its performance and areas of application.

Overview and positioning
Gemini 4 Carbon is Google’s most powerful model, designed for long-running agentic workflows such as cybersecurity, complex knowledge work, software engineering, and scientific reasoning. While Gemini 4 Argon is suited to everyday tasks, Carbon excels in situations requiring more tokens for thorough verification and deep reasoning. It also ranked first in independent benchmarks such as CursorBench 4.0 and Perplexity’s WANDR, and offers superior cost-performance compared with competing models.

Advanced knowledge work and data analysis
In fields such as law, finance, and corporate operations, it extracts accurate insights from large volumes of dense documents and data. It excels at generating spreadsheets, presentations, and reports, and has received top-tier ratings on the MKA benchmark, which measures multimodal performance. By using the new “/swarm” feature, multiple agents can collaborate from a single instruction to automatically perform tasks such as creating 3D exhibits from historical photographs and reconstructing buildings in 3D.

Agentic coding and UI/3D design
It supports advanced engineering tasks such as migrating multiple packages and redesigning architectures. It can complete web applications while self-repairing from errors and checking UI rendering, and demonstrates high comprehension and precision in areas such as 3D game development and physics simulation.

Enhanced cybersecurity capabilities
It has achieved top results in benchmarks for vulnerability discovery (SEC-bench Pro), exploitability validation (ExploitGym), and patch generation (CWE-bench v1). It has already delivered practical outcomes that support defenders, such as rapidly identifying zero-day vulnerabilities and fixing serious defects in open-source software.

Scientific and mathematical reasoning
It has produced results in advanced research support, including identifying toxicity mechanisms in drug discovery, formally verifying the Signal encryption protocol (using Lean code generated by AlphaProof Code), diagnosing experimental noise in enzyme assays, and conducting large-scale searches for homologous proteins for applications such as AlphaFold 3.

Safety and security measures
Alongside improvements in reasoning capabilities, measures to prevent misuse have also been strengthened. It blocks 99% of attack attempts through adaptive red teaming and records a 0% attack success rate on the indirect prompt injection benchmark (Gray Swan IPI). It also clears 92% of tasks on the agent honesty benchmark and substantially reduces data fabrication.

Availability
It is available through paid developer APIs (Google AI Studio, Android Studio) and Gemini for Business at the same price as Argon ($2 per million input tokens and $10 per million output tokens). It is also being gradually rolled out through the Gemini app, Google Antigravity, and other services for Google AI Ultra subscribers. A six-month free access program is also available for researchers.

Google DeepMind has announced a new framework, “AlphaProof Code,” that mathematically verifies Signal protocol implementation code using AI agents.

Based on expert-defined specifications, the AI generated more than 6 million lines of Lean code and produced a proof. It completed a verification that would ordinarily take an expert team years in just four weeks, demonstrating that the cryptographic protocol’s safety guarantees hold even in the compiled executable code.

Google DeepMind has released a new benchmark, “MKA,” that measures AI agents’ ability to interpret visually complex files and perform practical tasks. It consists of 200 real-world tasks across 10 fields and 83 occupations and was developed with experts from those fields.

According to the results, the new model “Gemini 4 Carbon” ranked second in quality while demonstrating the ability to complete tasks at an average cost of one-sixth that of the highest-scoring model.

MaxText has introduced context parallelism for Cloud TPUs.

It leverages four parallelization strategies and high-speed attention kernels to efficiently enable long-context training of up to 14 million tokens.

Anthropic

This guide explains prompting techniques specific to Claude Fable 5.2 and Mythos 5.2. Taking into account differences in behavior from earlier models, it provides optimal instructions for different situations.

Specifically, it explains how to adjust the “effort” setting, which determines processing capacity, and how to prevent tasks from stopping prematurely. It also compiles practical techniques, such as explicitly clarifying code ownership, appropriately delegating work to subagents, and avoiding excessive caution caused by unnecessary instructions from the past.

When migrating to a new model, it is recommended to apply these guidelines in order and test their effectiveness for each task.

A team of 1,000 agents autonomously hierarchicalized and divided tasks under shared database management and strict safety controls, reproducing a formal proof of Fermat’s Last Theorem in 28.7 hours.

Anthropic has begun offering Whiteboards, Sheets, and Claude Data, which enable collaborative work within Claude.

Users can brainstorm, work with spreadsheets, and share data directly without switching to external tools. These features are available with paid plans.

Meta

Meta has announced “Muse Ember,” a small AI model that reduces the cost of iterative processing by AI agents.It efficiently handles sequential tool calls, code modifications, and data integration between apps.

It supports a context window of 1 million tokens and can also handle images, audio, and video. In performance benchmarks, it achieves scores approaching those of Claude Haiku 5.5 at roughly one-third the cost, and it has also begun to be offered as an API.

Meta's personal AI agent, “Muse,” now supports Apple Watch and smartwatches running Google's Wear OS.

Muse is an AI that creates plans based on the user's preferences and carries out tasks such as making reservations. It has a mechanism that requests confirmation before sending emails or making purchases.

In Japan, it is being rolled out in phases from late October 2026, and with this support, users can now operate it directly from their wrist.

OpenAI

In OpenAI's personal AI agent, “dots,” you can now connect multiple local PCs to a single dot and select the PC to use for each task. While the user continues working on their laptop, app builds and tests can be run in parallel on another PC.

To use it, grant access through the ChatGPT app on the PC and instruct the dot which PC to target and what work to perform. GPT-6.1 Terra has also been added to the available models, improving both response speed and stability.

Papers & Other

The problem of AI agents executing tools or external actions twice due to communication drops or processing delays cannot be fully prevented through prompt engineering alone. It can be effectively prevented by appropriately pinning keys on the execution infrastructure side, ensuring idempotency (duplicate prevention) in the destination service, and asking a human to make a judgment when uncertain.

You do not need to address every point raised in an AI review of a document. It is effective to categorize the feedback as “fix,” “verify,” or “don't fix,” based on factors such as whether the premise is correct and whether it duplicates existing statements, and then ask for only the points that should be corrected in a new chat.

Engineers

“M3E Canvas,” open-source software developed by inori, a first-year university student, has attracted attention after gaining more than 10,000 GitHub stars in the roughly two months since its release. It originated from the challenge of “it being a hassle to give instructions about the screen” when developing apps with AI. By arranging UI components in a browser, it can automatically generate highly accurate instruction prompts to pass to AI.

It was initially intended for individual developers in Japan, but its intuitive operation and other features were highly praised, and it gained explosive popularity, especially overseas. Currently, it is used for a wide range of purposes beyond the author's expectations, including not only prompt generation but also product mockup creation and screen design for web and desktop apps other than Android.

The author prioritizes enjoyment and ease of use over personal profit and has indicated an intention to continue development while valuing the open-source culture.

Related

Cloud

Azure

The Azure DocumentDB team has announced that it will participate in Microsoft Ignite 2026 in San Francisco and hold a breakout session on migrating and modernizing MongoDB applications.

OS

macOS

Apple has released “iOS/iPadOS 18.7.12,” “iOS/iPadOS 26.7.3,” and “watchOS 26.6.1,” fixing multiple vulnerabilities for older iPhones and Apple Watches.

Chips

In multicore processors, work that cannot be parallelized becomes a bottleneck to performance improvements. This is illustrated by “Amdahl’s law.” By contrast, “Gustafson’s law” has also been proposed, asserting that increasing the scale of the tasks handled can improve parallelism.

Currently, a line graph known as the “Roofline model” is used as an easy-to-understand visualization of whether “memory bandwidth” or “computational performance” is the factor limiting multicore execution performance.

AMD’s Zen 6 family has evolved primarily at the front end, including faster branch prediction and the use of dual decoding on a single thread.

A major change is that Zen 6 and Zen 6c now have different architectures. While Zen 6 is tuned for high clock speeds, Zen 6c expands out-of-order structures such as the ROB to improve tolerance of memory latency. Design specialization for different use cases is advancing beyond the traditional differences in density.

Google has added the new “TPU 8tx” and “TPU 8i Max” systems to its eighth-generation TPU portfolio for the agent era.

The 8tx protects data at the hardware layer, enabling secure confidential inference and private tuning. The 8i Max uses a two-chip configuration to double computing power and memory bandwidth, eliminating the memory bottleneck in long-context processing.

Software capabilities, including PyTorch support, have also been enhanced.

Industry Trends and Current Events

Adobe and Disney have announced a multi-year strategic partnership focused on the use of AI tools.

They will build proprietary generative AI models based on Disney’s intellectual property. These models will be used in theme park development and other production environments to accelerate idea generation and work.

SoftBank Group’s consolidated financial results for the April–September 2026 period were affected by massive losses related to changes in its US AI data center plans, resulting in a final net loss of ¥768 billion.

The Bank of Japan’s December additional rate hike expectations, which temporarily declined amid market turmoil following the US midterm elections, have recovered to the 70 percent range, maintaining the view that the BOJ will raise rates for the third time this year—the first such hike in 37 years.

Mr. Schmidhuber, known as the “father of modern AI,” points out that current large language models are nothing more than imitations of existing data, and that achieving human-level intelligence (AGI) requires “physical AI” that understands the physical world.

The development of autonomous robots that learn physical laws on their own will be key going forward, and he says that Japan, with its strengths in robotics, also has an opportunity. On the other hand, he is skeptical about the realization of AGI within the next few years because mastering the physical domain takes time.

Mr. Tanba, president of “Noetra,” a new AI development company launched with investments from 44 domestic companies, including SoftBank, said the company aims to build a domestically developed AI foundation—a “world model” that understands physical properties.

The idea is to ensure data sovereignty from the standpoint of economic security and apply the lessons of the feature-phone era, when the country confined itself to manufacturing devices and missed out on revenue from services.

Although development is currently being advanced using Nvidia’s semiconductors, the company indicated that it will also explore alternatives to the company for practical deployment.