2025-12-16 Tech Briefing | Ant Group open-sources Agentar SQL for data intelligence.

Tuesday

Welcome to this edition, featuring 32 articles.

📚 Contents/📝 TL;DR

📚 Contents

📝 TL;DR

Another Data Analytics Intelligent Scale Opens Source, Now Ranking BIRD! Non-professionals Can Easily Explore and Analyze Data

Ant Group open-sources Agentar SQL for data intelligence.

Another Data Analytics Intelligent Scale Opens Source, Now Ranking BIRD! Non-professionals Can Easily Explore and Analyze Data

Ant Group open-sources Agentar SQL for data intelligence.

Using kro to Build Composable Platforms

CNCF explores how kro helps build Kubernetes composable platforms.

Using kro to Build Composable Platforms

CNCF explores how kro helps build Kubernetes composable platforms.

How Booking.com Built a Low-Latency Feature Platform Using Amazon ElastiCache

Booking.com uses ElastiCache to build a low-latency ML feature platform.

How Booking.com Built a Low-Latency Feature Platform Using Amazon ElastiCache

Booking.com uses ElastiCache to build a low-latency ML feature platform.

ChatGPT's Competitors: The Quiet Rise of Kuaishou: The Most Popular Internet Service of 2025

Cloudflare releases a report on the most popular internet services of 2025, revealing competitive landscapes in various fields

ChatGPT's Competitors: The Quiet Rise of Kuaishou: The Most Popular Internet Service of 2025

Cloudflare releases a report on the most popular internet services of 2025, revealing competitive landscapes in various fields

AI's "Brain Circuits" Finally Visible? OpenAI Releases Sparse Model to Make Inference Transparent

OpenAI's sparse model explores new paths for AI explainability

AI's "Brain Circuits" Finally Visible? OpenAI Releases Sparse Model to Make Inference Transparent

OpenAI's sparse model explores new paths for AI explainability

December 15, 2025

A quick overview of global headlines on December 15, 2025.

December 15, 2025

A quick overview of global headlines on December 15, 2025.

Amazon EKS Introduces Enhanced Network Policy Features

EKS introduces cluster-wide Admin Policies and Application Network Policies supporting FQDN-based egress filtering.

Amazon EKS Introduces Enhanced Network Policy Features

EKS introduces cluster-wide Admin Policies and Application Network Policies supporting FQDN-based egress filtering.

Luminar, the Laser Radar Giant, Files for Bankruptcy: 30 Billion Valuation Turns to Zero, Assets Less Than Liabilities, Founder Attempts to Acquire Former Team but is Rejected?

Luminar, the laser radar giant, files for bankruptcy with a 30 billion valuation collapsing, founder's attempt to acquire former team fails.

Luminar, the Laser Radar Giant, Files for Bankruptcy: 30 Billion Valuation Turns to Zero, Assets Less Than Liabilities, Founder Attempts to Acquire Former Team but is Rejected?

Luminar, the laser radar giant, files for bankruptcy with a 30 billion valuation collapsing, founder's attempt to acquire former team fails.

How to Create a SQL Server Developer Edition Instance on Amazon RDS for SQL Server

AWS RDS supports SQL Server Developer Edition, saving licensing fees.

How to Create a SQL Server Developer Edition Instance on Amazon RDS for SQL Server

AWS RDS supports SQL Server Developer Edition, saving licensing fees.

Vitest Team Releases 4.0 Version, Providing Stable Browser Mode and Visual Regression Testing Features

Vitest 4.0 offers stable browser mode and visual regression testing.

Vitest Team Releases 4.0 Version, Providing Stable Browser Mode and Visual Regression Testing Features

Vitest 4.0 offers stable browser mode and visual regression testing.

Kubernetes Security: Stable Features in 2025 and Preview Features in 2026

K8s security features are stable in 2025 and preview in 2026.

Kubernetes Security: Stable Features in 2025 and Preview Features in 2026

K8s security features are stable in 2025 and preview in 2026.

Essential

DeepSeek Forces vLLM to Upgrade! Chip Competition, MoE Sweeps Across Models, vLLM Core Maintainer Exclusive Response: How to Secure the Inference 'Throne' with PyTorch

vLLM leverages PyTorch and community power to face chip and MoE model challenges.

DeepSeek Forces vLLM to Upgrade! Chip Competition, MoE Sweeps Across Models, vLLM Core Maintainer Exclusive Response: How to Secure the Inference 'Throne' with PyTorch

vLLM leverages PyTorch and community power to face chip and MoE model challenges.

From 4 Billion Device Vulnerabilities to AI Browsers: 3 'Invisible' Principles in Browser Underlying

Deep analysis of browser underlying principles and frontend pitfalls.

From 4 Billion Device Vulnerabilities to AI Browsers: 3 'Invisible' Principles in Browser Underlying

Deep analysis of browser underlying principles and frontend pitfalls.

GPT-5.2 Makes a Full-Fledged Push! Crushes 44 Professional Tasks, Real-World Testing Shows No Rival in Programming at Same Price, Deep Reasoning Dominates, but Speed is a Major Weakness

OpenAI releases GPT-5.2, significantly enhances professional work capabilities, strong in programming and reasoning but slow in speed.

GPT-5.2 Makes a Full-Fledged Push! Crushes 44 Professional Tasks, Real-World Testing Shows No Rival in Programming at Same Price, Deep Reasoning Dominates, but Speed is a Major Weakness

OpenAI releases GPT-5.2, significantly enhances professional work capabilities, strong in programming and reasoning but slow in speed.

What AWS Security Team Learned from Recent npm Supply Chain Threat Activities

AWS shares experience in responding to npm supply chain attacks, emphasizing defense.

What AWS Security Team Learned from Recent npm Supply Chain Threat Activities

AWS shares experience in responding to npm supply chain attacks, emphasizing defense.

OpenAI Secretly Integrates Anthropic Skills, ChatGPT and Codex Now Carry the Mechanism! 111 Minutes to Generate PDF and Auto-Font Adjustment, Developers Call It More Important Than MCP!

OpenAI adopts Anthropic Skills mechanism; Agent capabilities are enhanced.

OpenAI Secretly Integrates Anthropic Skills, ChatGPT and Codex Now Carry the Mechanism! 111 Minutes to Generate PDF and Auto-Font Adjustment, Developers Call It More Important Than MCP!

OpenAI adopts Anthropic Skills mechanism; Agent capabilities are enhanced.

How Temporal Helps Netflix Achieve Reliable Cloud Operations

Netflix reduced deployment failure rate from 4% to 0.0001% by using Temporal.

How Temporal Helps Netflix Achieve Reliable Cloud Operations

Netflix reduced deployment failure rate from 4% to 0.0001% by using Temporal.

Amazon Threat Intelligence: Russian Cyber Threat Group Targets Western Critical Infrastructure

Amazon reveals a Russian APT group targeting Western critical infrastructure.

Amazon Threat Intelligence: Russian Cyber Threat Group Targets Western Critical Infrastructure

Amazon reveals a Russian APT group targeting Western critical infrastructure.

NVIDIA Dynamo Addresses the Challenges of Multi-Node Large Language Model Inference

NVIDIA Dynamo framework solves multi-node LLM inference challenges, improving efficiency.

NVIDIA Dynamo Addresses the Challenges of Multi-Node Large Language Model Inference

NVIDIA Dynamo framework solves multi-node LLM inference challenges, improving efficiency.

Top AI Agentic Workflow Patterns

This article introduces five mainstream AI Agent workflow patterns.

Top AI Agentic Workflow Patterns

This article introduces five mainstream AI Agent workflow patterns.

2025 Cloudflare Radar Annual Review: AI, the Post-Quantum Era, and the Rise of Record-Breaking DDoS Attacks

Cloudflare releases 2025 annual internet report, covering traffic, AI, and security trends

2025 Cloudflare Radar Annual Review: AI, the Post-Quantum Era, and the Rise of Record-Breaking DDoS Attacks

Cloudflare releases 2025 annual internet report, covering traffic, AI, and security trends

Is AI Coding Tool Becoming a 'Formatting God'? Claude CLI Has Become a 'System Killer' in Half a Year, Developers Accuse: All Their Work Was Lost!

Claude CLI mistakenly executed the `rm -rf ~/` command, leading to data loss for multiple developers.

Is AI Coding Tool Becoming a 'Formatting God'? Claude CLI Has Become a 'System Killer' in Half a Year, Developers Accuse: All Their Work Was Lost!

Claude CLI mistakenly executed the `rm -rf ~/` command, leading to data loss for multiple developers.

Kubernetes Security: Stable Features in 2025 and Preview Features in 2026

K8s security features are stable in 2025 and preview in 2026.

Kubernetes Security: Stable Features in 2025 and Preview Features in 2026

K8s security features are stable in 2025 and preview in 2026.

Amazon Web Services Launches Persistent Functions, Enabling Stateful Logic in Lambda Code

AWS Lambda supports stateful persistent functions.

Amazon Web Services Launches Persistent Functions, Enabling Stateful Logic in Lambda Code

AWS Lambda supports stateful persistent functions.

TPU Orders Surge, Google Expands Production of Next-Gen Chips! Google's Chief Scientist: We've Used It for Over a Decade and Have Been Very Satisfied

Google's TPU orders surge, Jeff Dean discusses the 10-year vertical integration strategy.

TPU Orders Surge, Google Expands Production of Next-Gen Chips! Google's Chief Scientist: We've Used It for Over a Decade and Have Been Very Satisfied

Google's TPU orders surge, Jeff Dean discusses the 10-year vertical integration strategy.

Snippet: December 16th

AI programming paradigm shift: from writing code to designing closed-loop systems.

Snippet: December 16th

AI programming paradigm shift: from writing code to designing closed-loop systems.

General

Building Smarter Workflows with Notion and v0

v0 now integrates with Notion knowledge base, enhancing development workflows.

Building Smarter Workflows with Notion and v0

v0 now integrates with Notion knowledge base, enhancing development workflows.

React Beginner's Guide: Build Web Pages Like Building Blocks, JSX Makes Development a Joy!

React basics tutorial covering JSX, componentization, and environment setup.

React Beginner's Guide: Build Web Pages Like Building Blocks, JSX Makes Development a Joy!

React basics tutorial covering JSX, componentization, and environment setup.

Cultivating the Mindset of a Data Superhero

Snowflake's Data Superhero community recognition program is now open for applications.

Cultivating the Mindset of a Data Superhero

Snowflake's Data Superhero community recognition program is now open for applications.

Writing Fragments

Fowler reclaims RSS blogs due to social media fragmentation.

Writing Fragments

Fowler reclaims RSS blogs due to social media fragmentation.

2025 China Enterprise AI Office Digitalization White Paper

Overview of the "2025 China Enterprise AI Office Digitalization White Paper".

2025 China Enterprise AI Office Digitalization White Paper

Overview of the "2025 China Enterprise AI Office Digitalization White Paper".

re:Invent Special Planning: 'Five Days, Five Questions' ⑤ 'Forgetting and Action' Launches!

The fifth episode of re:Invent review video is now online.

re:Invent Special Planning: 'Five Days, Five Questions' ⑤ 'Forgetting and Action' Launches!

The fifth episode of re:Invent review video is now online.

Primary

Another Data Analytics Intelligent Scale Opens Source, Now Ranking BIRD! Non-professionals Can Easily Explore and Analyze Data

InfoQ - AI & LLM•Published on December 15

Insight•AI & Emerging Technologies•Score: 8/10

Data Intelligent AgentNL2SQLAgentar SQLAnt GroupOpen Source

Source

◆

Using kro to Build Composable Platforms

CNCF•Published on December 15

Insight•Cloud Computing & Backend•Score: 8/10

kroKubernetesPlatform EngineeringResource CompositionComposition

📈

Perspectives

As of December 2025, platform engineering in China has moved from concept to implementation, with intense competition. CNCF's promotion of projects like kro reflects the industry's demand for standardized, portable platform building foundations, especially in the context of multi-cloud and hybrid cloud trends, effectively reducing vendor lock-in risks. However, for domestic enterprises, the challenge lies more in aligning business complexity with team scale than in pure technical composition. Tools like kro can address technical 'composition' issues, but the real test for architects and product managers is how to integrate technical capabilities with enterprise-specific operational processes, compliance requirements, and legacy systems. Don't expect a single kro to solve everything.

Source

◆

How Booking.com Built a Low-Latency Feature Platform Using Amazon ElastiCache

AWS - Databases•Published on December 15

Insight•AI and Frontier Technologies•Score: 8/10

feature platformElastiCachelow latencymachine learningcost optimization

🔬

Tech Insights

Booking.com's case highlights: 1. 'Extreme performance and scale': achieved p99.9 latency under 25 milliseconds, handling 200,000 requests per second (RPS) while ingesting 50,000 RPS. These performance metrics are top-tier globally, demonstrating ElastiCache's effectiveness in specific scenarios. 2. 'Cost-effective': migrating to the ElastiCache-based feature platform reduced infrastructure costs by 82% and saved over $3 million annually. This is the most eye-catching part, as performance is a bonus, while 'cost savings are the real deal'. 3. 'Self-service and workload isolation design': by centralizing schema and configuration management in a Git repository, CI/CD automation deployment is achieved, significantly reducing the operational burden on the core team. Each use case deploys an independent ElastiCache cluster, enabling workload isolation and resource customization, which is ideal for large organizations with multi-team collaboration.

Source

◆

ChatGPT's Competitors: The Quiet Rise of Kuaishou: The Most Popular Internet Service of 2025

Cloudflare Blog•Published on December 15

News•General News•Score: 7/10

Source

◆

AI's "Brain Circuits" Finally Visible? OpenAI Releases Sparse Model to Make Inference Transparent

InfoQ - AI & LLM•Published on December 15

Insight•AI and Frontier Technologies•Score: 7/10

AI ExplainabilitySparse ModelOpenAITransformerModel Biology

⚠️

Notable

The article also candidly points out that the current sparse Transformer has its limitations: training and inference speeds are 100 to 1000 times slower than dense models, due to the fact that current hardware (GPU/TPU) is optimized for dense matrices. This indicates that it is not feasible to directly apply to cutting-edge large models in the short term. Although the model structure is transparent, manual interpretation of the "circuit diagram" is still time-consuming, and there remains a gap between "understanding" and "explaining". Rather than saying "seeing the brain circuits", it is more like "designing a simpler brain circuit for you to see more easily", which is a strategy to simplify the problem space. In the short term, it has no direct architectural or product implementation value.

Source

◆

December 15, 2025

Global Vision•Published on December 15

News•Comprehensive News•Score: 7/10

Global NewsEconomic DataInternational RelationsAI RegulationClimate Change

Source

◆

Amazon EKS Introduces Enhanced Network Policy Features

AWS - Containers•Published on December 15

News•Cloud Computing and Backend•Score: 7/10

EKSnetwork policyAdmin PolicyFQDNsecurity

🔬

Tech Insights

1. Admin Network Policies: This is a boon for enterprise security compliance! Previously, Kubernetes' default network policies allowed all traffic and only operated at the Namespace level, making it difficult for cluster administrators to enforce consistent security policies. Now, with the Admin Tier, administrators can define immutable cluster-level security rules (e.g., 'block external internet access'), which is the dream of platform and security teams for centralized control. 2. Tiered Policy Priority: The evaluation order of Admin Tier > Standard Network Policies > Baseline Tier clearly defines the priority and override relationships of policies, making security policy creation and enforcement more structured and predictable. This resolves conflicts and chaos during multi-team collaboration. 3. Application Network Policies (FQDN Filtering): This feature is a 'game-changer' for traditional IP/CIDR-based egress policies. IPs for SaaS services or internal APIs often change, making IP whitelisting a tedious and error-prone task. Now, using FQDN directly enhances policy stability and maintainability, especially beneficial for microservices architectures that frequently access external services. 4. EKS Auto Mode Exclusive Feature: FQDN filtering is currently only available for EKS Auto Mode clusters, indicating that AWS is gradually guiding users toward more managed and automated EKS modes, reducing operational burden.

📈

Perspectives

1. While domestic cloud vendors have largely provided similar foundational capabilities in Kubernetes network policies, AWS's implementation of cluster-level, tiered, and FQDN-supporting policies still holds a competitive edge. Especially in the context of increasingly stringent enterprise-level security and compliance requirements, this fine-grained management capability is a core competitive advantage. 2. FQDN filtering is expected to become a standard feature for Kubernetes egress traffic management. Domestic cloud vendors and open-source network plugins (such as Calico, Cilium) will inevitably follow, potentially offering even richer Layer 7 policy features. 3. The limitation to EKS Auto Mode also signals a trend toward managed services. Enterprise users are increasingly inclined to delegate infrastructure complexity to cloud providers and focus on business logic. In the future, cloud-native products that offer higher-level abstraction and automated management will be more market-preferred.

Source

◆

Luminar, the Laser Radar Giant, Files for Bankruptcy: 30 Billion Valuation Turns to Zero, Assets Less Than Liabilities, Founder Attempts to Acquire Former Team but is Rejected?

InfoQ - AI & LLM•Published on December 16

News•Comprehensive News•Score: 7/10

Laser RadarLuminarBankruptcyAutonomous DrivingValuation Collapse

Source

◆

How to Create a SQL Server Developer Edition Instance on Amazon RDS for SQL Server

AWS - Databases•Published on December 15

News•Cloud Computing and Backend•Score: 7/10

AWS RDSSQL ServerDeveloper EditionCEVDatabase

Source

◆

Vitest Team Releases 4.0 Version, Providing Stable Browser Mode and Visual Regression Testing Features

InfoQ - Frontend•Published on December 16

News•Front-End Development•Score: 7/10

VitestFront-End TestingBrowser ModeVisual Regression TestingVite

🔬

Tech Insights

The stabilization of the browser mode is a major upgrade. Previously, front-end testing was often concerned with the differences between jsdom and real browser environments. Now, tests can run in real browsers, yielding more reliable results. For component libraries and framework development teams that heavily rely on DOM operations and browser APIs, this is a godsend. The integration of visual regression testing and Playwright Traces is also impressive, as it directly incorporates the functionalities of external tools like Percy/Chromatic, further simplifying the testing toolchain. This indicates that Vitest is evolving into a 'feature-complete testing platform,' aiming to replace Jest's dominance in the front-end field, and it's coming with the 'native speed advantages' of Vite.

Source

◆

Kubernetes Security: Stable Features in 2025 and Preview Features in 2026

InfoQ - Backend•Published on December 16

Insight•Cloud Computing and Backend•Score: 7/10

Kubernetes SecurityDevSecOpsCloud NativeK8sContainer Security

Source

Essential

DeepSeek Forces vLLM to Upgrade! Chip Competition, MoE Sweeps Across Models, vLLM Core Maintainer Exclusive Response: How to Secure the Inference 'Throne' with PyTorch

InfoQ - AI & LLM•Published on December 15

Insight•AI and Frontier Technologies•Score: 10/10

vLLMLarge Model InferencePyTorchHardware CompatibilityMoE

🔬

Tech Insights

1. PyTorch as the 'GCD': vLLM selects PyTorch as the core abstraction layer connecting upper-level frameworks with lower-level hardware. This is a brilliant move. As long as hardware vendors support PyTorch, vLLM can quickly adapt, significantly reducing the complexity of multi-hardware support, and explains why PyTorch has become the de facto standard in AI. 2. DeepSeek's Pressure and MoE Generalization: The release of DeepSeek R1, especially its MoE architecture, forced the vLLM team to rapidly iterate, transforming previously proprietary technologies into general implementations. vLLM thus evolved from supporting tensor parallelism and pipeline parallelism to supporting expert parallelism, which is crucial for handling trillion-parameter models. 3. Deep Collaboration with Hardware Ecosystems: The vLLM team collaborates deeply with NVIDIA, AMD, Google TPU, and even domestic chip manufacturers like MXChip. This joint effort advances hardware support and performance optimization. Red Hat's acquisition of Neural Magic and its active involvement in vLLM also underscores the critical role of the open-source community in AI infrastructure. 4. Multi-modal Prefix Caching and Encoder Decoupling: The release of vLLM-Omni introduced multi-modal prefix caching and encoder decoupling, transforming vLLM from a pure text inference engine into a unified multi-modal service platform. This directly addresses the efficiency and scalability issues in multi-modal model inference, marking a necessary step for future AI application development. 5. Community-Driven 'Speed Advantage': vLLM collaborates with numerous model labs and companies to rapidly gather feedback and feed improvements back into the community, creating a virtuous cycle. This extreme iteration speed makes private branches difficult to keep up, ultimately propelling vLLM to become the de facto standard for large model inference.

📈

Perspectives

1. vLLM is nearly an infrastructure-level presence in the domestic large model inference market, widely used by both big companies and startups. Its support for multi-hardware, especially its proactive embrace of domestic chip manufacturers, gives it strategic value in the wave of domestic substitution and self-reliance. 2. 'Chip competition' is a real situation in the domestic AI hardware market. The multi-hardware compatibility strategy of vLLM via PyTorch abstraction layer is the best path for numerous NPU and GPU vendors in China to enter the large model ecosystem. 3. Multi-modal is the future, and the release of vLLM-Omni indicates that it will become the core foundation for building multi-modal AI applications. In the future, the ecosystem around vLLM will become increasingly rich, including model deployment, fine-tuning, and agents, making it the 'web server' of the AI era. 4. Red Hat's involvement is expected to accelerate the enterprise-level adoption and standardization of vLLM. This is a major benefit for companies wanting to bring large model capabilities into their internal systems but worried about the maintenance and support of open-source projects. However, it also raises concerns about potential community fragmentation or strategic shifts due to commercialization.

Source

◆

From 4 Billion Device Vulnerabilities to AI Browsers: 3 'Invisible' Principles in Browser Underlying

Juejin Frontend Hottest•Published on December 15

Insight•Frontend Development•Score: 9/10

Browser UnderlyingWebXR VulnerabilityProcess IsolationIncremental RenderingAI Browser

Source

◆

GPT-5.2 Makes a Full-Fledged Push! Crushes 44 Professional Tasks, Real-World Testing Shows No Rival in Programming at Same Price, Deep Reasoning Dominates, but Speed is a Major Weakness

InfoQ - AI & LLM•Published on December 16

Insight•AI and Frontier Technologies•Score: 9/10

GPT-5.2OpenAIlarge modelsAI programmingdeep reasoning

Source

◆

What AWS Security Team Learned from Recent npm Supply Chain Threat Activities

AWS - Security•Published on December 15

Insight•Software Architecture and Engineering Practices•Score: 9/10

supply chain securitynpm securitysoftware securityDevSecOpsthreat intelligence

Source

◆

OpenAI Secretly Integrates Anthropic Skills, ChatGPT and Codex Now Carry the Mechanism! 111 Minutes to Generate PDF and Auto-Font Adjustment, Developers Call It More Important Than MCP!

InfoQ - AI & LLM•Published on December 16

Insight•AI and Frontier Technologies•Score: 9/10

OpenAI SkillsAnthropic SkillsAgentic AIMCPLLM tools

📈

Perspectives

OpenAI's embrace of Skills is essentially catching up, addressing Anthropic's early lead in Agent architecture. This indicates that the competition in Agentic AI is entering the 'capability encapsulation and management' phase. The core competitiveness of future AI products and services will shift from 'having the strongest base model' to 'building the most powerful and flexible Skills ecosystem.' Whoever makes it easier for enterprise users and developers to create, share, and combine Skills will gain an advantage in the enterprise AI market. Traditional application markets may be disrupted by this new 'Skills market' paradigm. In 2025, Chinese large model vendors that are still obsessed with 'parameter competition' and 'hallucination optimization' but ignore the idea of abstracting capabilities into manageable assets are not just 'late,' they are 'off track.'

Source

◆

How Temporal Helps Netflix Achieve Reliable Cloud Operations

Netflix Tech•Published on December 15

Insight•Software Architecture and Engineering Practices•Score: 9/10

TemporalNetflixSpinnakerReliabilityDistributed Systems

🔬

Tech Insights

1. Core Value: As a "persistent execution platform," Temporal allows Netflix to write code as if failures do not exist, which is a game-changer for handling transient failures in distributed systems (network, service interruptions, etc.). It abstracts away retries, timeouts, and state management, making business logic purer. 2. Pain Point Resolution: The original Spinnaker's Orca and Clouddriver introduced significant "undifferentiated complexity" to achieve reliability, such as building custom orchestration engines, complex retry logic, Saga framework, and local instance task states, which led to single points of failure. The introduction of Temporal directly replaced these redundant and fragile custom components. 3. Architecture Decoupling: By using Temporal as an intermediate layer, Orca no longer needs to engage in complex polling and state synchronization with Clouddriver, greatly reducing service coupling and enhancing maintainability. 4. Enforced Idempotency: Migrating cloud operation steps to Temporal Activities enforces logical idempotency, inherently solving potential issues caused by retries in Clouddriver previously. 5. Observability: Temporal's built-in UI and metrics make debugging and tracing workflow execution exceptionally simple, which is invaluable in complex production environments.

📈

Perspectives

1. Temporal, as a persistent workflow engine, is gradually gaining attention in the domestic market, especially in industries such as finance and e-commerce that have extremely high requirements for transaction consistency and long process reliability. In the past, we often talked about "business process orchestration," but platforms that can truly achieve self-healing, state persistence, and traceability are rare. 2. While there are traditional BPMN engines like Activiti and Camunda, Temporal is more developer-friendly and code-driven, aligning better with modern microservices architecture practices. Its "code as workflow" philosophy tightly integrates business logic with orchestration logic, reducing the gap between models and code. 3. Netflix's case is undoubtedly a strong endorsement for Temporal, proving its practical value in massive-scale, high-availability scenarios. It is expected that more domestic large companies will follow, and even cloud services or SaaS products based on Temporal may emerge to lower its adoption threshold. However, for most companies, building their own Temporal platform remains a challenge.

Source

◆

Amazon Threat Intelligence: Russian Cyber Threat Group Targets Western Critical Infrastructure

AWS - Security•Published on December 15

News•Cloud Computing and Backend•Score: 9/10

cybersecuritycritical infrastructureRussian APTcredential theftAWS defense

🔬

Tech Insights

The value of this report lies in its revelation of current cybersecurity attack and defense trends and high-risk areas: - Tactic Evolution is Worrying: Attackers have shifted from traditional 0-day/N-day exploit-based methods to using 'misconfigured network edge devices' as an initial access vector. This indicates that even state-sponsored APT groups are adopting a 'low-cost, high-efficiency' approach, presenting a greater challenge to many enterprises' basic security defenses. - Attack Targets and Chain are Clear: The group continues to focus on Western critical infrastructure, particularly the energy sector. The attack chain is: infiltrate edge devices -> leverage native packet capture capabilities -> steal credentials -> replay credentials to attack online services -> establish persistent access and lateral movement. This pattern is highly reusable and deserves caution from all enterprises. - Practical Defensive Recommendations: The report provides very specific defensive measures, such as immediately auditing network edge devices, enhancing credential management (MFA, anomaly detection), isolating management interfaces, and enabling AWS security services (GuardDuty, CloudTrail, Inspector). These are actionable steps and not just theoretical discussions.

⚠️

Notable

Although the report is comprehensive, as a senior architect, I can identify a few shortcomings: - The report, while detailed, could provide more specific definitions and common scenarios of 'misconfiguration' to better guide users in self-audits, rather than general advice. - The IOCs (Indicators of Compromise) provided are useful, but the emphasis that they 'may serve multiple purposes' and 'should not be automatically blocked' introduces some operational ambiguity for security teams. More precise contextual information is needed to avoid false positives or missed alerts. - Regarding credential replay attacks, the report mentions detection but does not delve into how to prevent them at the source (such as enforcing hardware MFA or passwordless authentication). This is a direction worth exploring, as prevention is better than cure.

📈

Perspectives

This report serves as a warning: even top-tier APT groups are now turning to 'low-hanging fruit'—that is, leveraging user misconfigurations and security oversights rather than solely pursuing 0-day vulnerabilities. This indicates that the focus of cybersecurity defense is shifting from pure technical confrontation to a greater emphasis on basic security practices and operations management. Domestic enterprises, especially those involved in critical information infrastructure, must recognize that 'configuration security is core security' and not place all hopes on advanced protective technologies. Additionally, state-backed attacks have become normalized, so enterprises need to establish more robust threat intelligence sharing mechanisms and incident response processes. This is not merely a technical issue, but a national security concern. Domestic vendors must demonstrate stronger awareness and action in this area.

Source

◆

NVIDIA Dynamo Addresses the Challenges of Multi-Node Large Language Model Inference

InfoQ - AI & LLM•Published on December 16

Insight•AI and Frontier Technologies•Score: 9/10

NVIDIA DynamoLLM InferenceMulti-NodeGPU SchedulingKubernetes

📈

Perspectives

In my view, NVIDIA's release of Dynamo is another major move in its AI software ecosystem, aiming to solidify its position as the 'full-stack champion' in the AI field. It's not just about selling GPUs, but also about selling solutions, which is a smart commercial strategy. As large model applications become more widespread, inference costs have become the final straw for many organizations. Dynamo directly targets this pain point and has the potential to become the 'new cornerstone' of LLM inference infrastructure. For domestic AI companies, especially those heavily investing in large model applications, Dynamo offers a textbook-level optimization approach. If fully understood and localized, it can significantly enhance market competitiveness. This tool's release will likely make teams still using single-machine single-GPU inference tools feel like crying in the bathroom.

Source

◆

Top AI Agentic Workflow Patterns

ByteByteGo•Published on December 15

Insight•AI & Frontier Technologies•Score: 8/10

AI AgentAgentic WorkflowLLMReActMulti-agent

🔬

Tech Insights

In my view, the article's highlight is its clear summary of several Agentic workflow patterns: 1. Reflection Pattern (Reflective Pattern): Allows Agents to self-review and improve, which is essentially the AI version of traditional software engineering practices like 'test-driven development' or 'code review', representing a core method for improving output quality. 2. Tool Use Pattern (Tool Utilization Pattern): Grants Agents the ability to call external tools, marking a qualitative leap from LLMs being 'good at talking' to 'good at doing'. Without tools, LLMs are just talkers. 3. ReAct Pattern (Reason and Act): Integrates reasoning and action, simulating the human process of solving complex problems. This is more adaptive than pure planning or blind execution and serves as the foundation for complex tasks. 4. Planning Pattern (Planning Pattern): Emphasizes strategic thinking and task decomposition, which is crucial for multi-step, long-chain tasks to avoid Agents acting like aimless flies. 5. Multi-Agent Pattern (Multi-Agent Collaboration): Enables specialized division of labor among multiple Agents to complete tasks together, which is undoubtedly the ultimate form of building large-scale, complex AI systems, akin to a mini AI company. Additionally, these patterns are not mutually exclusive; real-world engineering often involves a combination of multiple patterns. Understanding these patterns is essential to avoid treating LLMs as a 'black box' and instead see them as a programmable, tunable intelligent unit.

Source

◆

2025 Cloudflare Radar Annual Review: AI, the Post-Quantum Era, and the Rise of Record-Breaking DDoS Attacks

Cloudflare Blog•Published on December 15

News•Comprehensive News•Score: 8/10

Cloudflare RadarInternet TrendsAI TrafficDDoS AttacksPost-Quantum EncryptionIPv6Zombie NetworksEmail Security2025 Annual Review

🔬

Tech Insights

This report contains massive data, and as an architect, I identified several key takeaways: 1. Exponential Growth of AI Traffic: AI crawlers such as Googlebot and OpenAI's GPTBot accounted for a significant portion of traffic, particularly 'user behavior crawling,' which increased over 15 times. This clearly indicates that AI is reshaping internet traffic patterns, and content owners must address the impact of AI scraping on bandwidth and content ownership. 2. Rapid Adoption of Post-Quantum Encryption: In 2025, the share of TLS 1.3 traffic supporting post-quantum encryption reached 52%. This shows that the threat of quantum computing is no longer a distant future but a present reality. Enterprises need to start planning their encryption strategies to defend against 'capture now, decrypt later' attack models. 3. Massive Scale of DDoS Attacks: The report highlights that peak DDoS attack sizes have reached 31.4 Tbps, with both frequency and scale continuing to rise. This underscores that DDoS defense remains a top priority in cybersecurity, especially for core business systems. 4. Slow but Steady IPv6 Adoption: Global IPv6 traffic share is only 29%, with India leading the way. This indicates that while the IPv4 exhaustion crisis persists, IPv6 deployment is still a long-term challenge, both domestically and internationally. 5. Zombie Networks and Abuse of Cloud Services: 40% of zombie network traffic originates from the U.S., with AWS and Google Cloud contributing a quarter of it. This suggests that cloud services, while convenient, have become a breeding ground for attackers, making cloud security governance urgent.

📈

Perspectives

In the context of the domestic market: * AI Traffic: The fierce competition among large AI models in China means that demand for data scraping during model training and RAG (retrieval-augmented generation) will only grow. Content providers must actively manage robots.txt and consider implementing AI crawler control mechanisms to protect their content value. * Cybersecurity: The trend of large-scale DDoS attacks is equally evident in China, especially in high-value sectors such as gaming and finance. Enterprises must continue to invest in DDoS defense and consider adopting multi-layered protection strategies. * IPv6 and Mobile Priority: Although IPv6 adoption in China is increasing, it has not yet reached an ideal level. Simultaneously, mobile device traffic accounts for more than half of the total, further confirming that the 'mobile-first' development and operation strategy is an irreversible trend. Any product targeting consumers (to C) that fails to deliver a good mobile experience is essentially 'goodbye.'

Source

◆

Is AI Coding Tool Becoming a 'Formatting God'? Claude CLI Has Become a 'System Killer' in Half a Year, Developers Accuse: All Their Work Was Lost!

InfoQ - AI & LLM•Published on December 16

News•Software Architecture & Engineering Practices•Score: 8/10

Claude CLIAI Coding ToolDatabase DeletionSecurity RiskAgent

Source

◆

Kubernetes Security: Stable Features in 2025 and Preview Features in 2026

CNCF•Published on December 15

Insight•Cloud Computing and Backend•Score: 8/10

Kubernetes SecurityDevSecOpsCloud NativeK8sContainer Security

Source

◆

Amazon Web Services Launches Persistent Functions, Enabling Stateful Logic in Lambda Code

InfoQ - Backend•Published on December 16

News•Cloud Computing & Backend•Score: 8/10

AWS LambdaPersistent FunctionsServerlessState ManagementWorkflow Orchestration

🔬

Tech Insights

This feature is a 'game-changer' for AWS Serverless architecture! Previously, Lambda functions were stateless, and to handle complex workflows, developers either had to manually implement state management or use Step Functions. Now, using context.step() and context.wait() within Lambda functions, developers can implement 'stateful logic' with 'state and computation separation,' which is the 'serverless orchestration of 'what you see is what you get.' The most impressive part is the ability to pause execution for up to a year without incurring costs, which greatly reduces costs and complexity for long-cycle approval flows, human intervention processes, or tasks like LLM that require extended processing time. It not only solves the 15-minute Lambda execution time limit but also allows developers to build workflows using more familiar programming languages.

📈

Perspectives

AWS's move is to 'erode part of Step Functions' market, but more importantly, it offers Serverless developers a 'more flexible and cost-effective' orchestration option. It greatly expands the applicability of Lambda, especially for teams that want to build complex workflows with pure code but are reluctant to be bound by the additional complexity (and cost) of Step Functions. In China, although Serverless architecture adoption is not as widespread as overseas, under the 'cost reduction and efficiency improvement' trend, 'Lambda persistent functions' as a 'cost-saving tool' will definitely gain popularity. Teams that previously hesitated due to Step Functions being too heavy now have a lighter alternative. This will also push domestic cloud vendors to consider similar 'stateful function' capabilities in their serverless computing services. In the long run, this is a signal that makes Serverless more 'versatile,' accelerating its adoption in China.

Source

◆

TPU Orders Surge, Google Expands Production of Next-Gen Chips! Google's Chief Scientist: We've Used It for Over a Decade and Have Been Very Satisfied

InfoQ - AI & LLM•Published on December 16

Insight•AI and Frontier Technologies•Score: 8/10

TPUGoogle DeepMindAI ChipsVertical IntegrationLarge Model Inference

Source

◆

Snippet: December 16th

Martin Fowler•Published on December 16

Insight•AI-Assisted Engineering•Score: 8/10

Agentic LoopCode ReviewHost Modernization

Source

General

Building Smarter Workflows with Notion and v0

Vercel News•Published on December 15

News•Software Architecture & Engineering Practices•Score: 6/10

Notionv0AI Code GenerationWorkflowDocument-Driven Development

Source

◆

React Beginner's Guide: Build Web Pages Like Building Blocks, JSX Makes Development a Joy!

Juejin Frontend Hottest•Published on December 13

Insight•Frontend Development•Score: 5/10

ReactJSXComponentizationFrontend FrameworkVite

🔬

Tech Insights

The article uses the metaphor 'building a web page like building blocks' to describe component-based development, which is very illustrative and indeed reflects the core concept of React and modern frontend frameworks. Breaking down complex pages into independent, reusable components is the foundation for improving development efficiency and code maintainability. It mentions both create-react-app and Vite as project setup methods and explicitly recommends Vite. In my opinion, teams still using CRA should consider upgrading unless they are bound by legacy systems, as Vite offers real improvements in development experience. The explanation of JSX emphasizes the rule that 'only expressions can be placed, not statements,' and provides code examples in common scenarios such as list rendering and conditional rendering. This helps newcomers understand the essence and boundaries of JSX.

📈

Perspectives

Although the content is basic, its status as 'This Week's Hot' reflects a reality: the flow of new talent into the frontend field remains strong, and the demand for foundational knowledge is huge. For training institutions and junior recruitment markets, React is still a must-have skill. From an industry development perspective, the 'entry barrier' for frontend frameworks is gradually decreasing, but the comprehensive ability requirements for building high-performance, high-maintainability complex applications are increasing. Simply mastering these basics is not enough. In the coming years, the competition in frontend frameworks will shift from 'basic syntax' to 'ecosystem maturity', 'development efficiency toolchains', and 'cross-platform/full-stack integration capabilities'. If one only stays at the level of 'JSX is fun', they will quickly be left behind by the market.

Source

◆

Cultivating the Mindset of a Data Superhero

InfoQ - Cloud•Published on December 16

Insight•Comprehensive News•Score: 5/10

SnowflakeCommunity OperationsKOLEcosystem BuildingPersonal Brand

Source

◆

Writing Fragments

Martin Fowler•Published on December 16

News•Content Distribution•Score: 5/10

RSSSocial MediaDecentralization

Source

◆

2025 China Enterprise AI Office Digitalization White Paper

InfoQ - AI & LLM•Published on December 15

News•Comprehensive Information•Score: 3/10

AI officedigital transformationwhite paperenterprise applicationsmarket trends

Source

◆

re:Invent Special Planning: 'Five Days, Five Questions' ⑤ 'Forgetting and Action' Launches!

InfoQ - AI & LLM•Published on December 15

News•Mixed News•Score: 1/10

re:InventAWSConference ReviewVideo