2026-02-01 Tech Briefing | Baidu opensources a new generation of OCR model, outperforming mainstream competitors.

Sunday

Welcome to this edition, featuring 21 articles.

📚 Contents/📝 TL;DR

📚 Contents

📝 TL;DR

Primary

OCR Competition Intensifies! Baidu Releases New SOTA OCR Model, Outperforms DeepSeek-OCR2

Baidu opensources a new generation of OCR model, outperforming mainstream competitors.

OCR Competition Intensifies! Baidu Releases New SOTA OCR Model, Outperforms DeepSeek-OCR2

Baidu opensources a new generation of OCR model, outperforming mainstream competitors.

Memory Prices Surge 170% Year-over-Year, It's Time to Clarify the 'Hidden Costs' in Cloud Bills

Memory prices have skyrocketed, and Huawei Cloud introduces flexible computing to optimize costs.

Memory Prices Surge 170% Year-over-Year, It's Time to Clarify the 'Hidden Costs' in Cloud Bills

Memory prices have skyrocketed, and Huawei Cloud introduces flexible computing to optimize costs.

Comparison between HTTP/2 Based on TCP and HTTP/3 Based on QUIC

ByteByteGo covers multiple topics, AI predictions, and system design reviews.

Comparison between HTTP/2 Based on TCP and HTTP/3 Based on QUIC

ByteByteGo covers multiple topics, AI predictions, and system design reviews.

AWS Releases Digital Sovereignty Lens for Well-Architected Framework

AWS releases a digital sovereignty architecture lens to guide users in building and operating compliant workloads in the cloud.

AWS Releases Digital Sovereignty Lens for Well-Architected Framework

AWS releases a digital sovereignty architecture lens to guide users in building and operating compliant workloads in the cloud.

Why Is There a Video Tag? Why Do We Still Need a Video Player?

Native video tag is insufficient; professional video player libraries are necessary.

Why Is There a Video Tag? Why Do We Still Need a Video Player?

Native video tag is insufficient; professional video player libraries are necessary.

Efficient JSON Data Analysis Using Amazon Redshift SUPER

Amazon Redshift SUPER data type enables efficient analysis of JSON data.

Efficient JSON Data Analysis Using Amazon Redshift SUPER

Amazon Redshift SUPER data type enables efficient analysis of JSON data.

Vercel Sandbox is Now Fully Available, Supporting Execution of Untrusted Code

Vercel Sandbox is now fully available, providing a secure and fast sandbox environment for AI Agents.

Vercel Sandbox is Now Fully Available, Supporting Execution of Untrusted Code

Vercel Sandbox is now fully available, providing a secure and fast sandbox environment for AI Agents.

Aspire 13.1 Release: Bringing MCP Integration, CLI Enhancements, and Azure Deployment Updates

Aspire 13.1 update enhances AI-assisted development, improves Azure deployment.

Aspire 13.1 Release: Bringing MCP Integration, CLI Enhancements, and Azure Deployment Updates

Aspire 13.1 update enhances AI-assisted development, improves Azure deployment.

Limited Budget, Technical Gaps: The Most Challenging AI Users Are Small and Medium Enterprise Owners

Huawei Cloud launches Flexus AI agent, targeting AI implementation for small and medium enterprises.

Limited Budget, Technical Gaps: The Most Challenging AI Users Are Small and Medium Enterprise Owners

Huawei Cloud launches Flexus AI agent, targeting AI implementation for small and medium enterprises.

Essential

Zhiyuan Multimodal Large Model Published in Nature, Establishing Autoregressive as a Unified Path for Generative AI

Google's World Model Project Genie Goes Viral: 'Splitting Wood' Brother and Hamersley Take the Stage, Revealing the 60-Second Limit is Not a Limit, Memory is the Major Constraint

Google releases the world model Genie, enabling real-time, interactive virtual worlds.

Google's World Model Project Genie Goes Viral: 'Splitting Wood' Brother and Hamersley Take the Stage, Revealing the 60-Second Limit is Not a Limit, Memory is the Major Constraint

Google releases the world model Genie, enabling real-time, interactive virtual worlds.

‘The World Has Suffered from CUDA for Too Long!’ KernelCAT Takes the Lead, Enabling Seamless Adaptation for Domestic Chips

KernelCAT achieves intelligent operator optimization for domestic chips.

‘The World Has Suffered from CUDA for Too Long!’ KernelCAT Takes the Lead, Enabling Seamless Adaptation for Domestic Chips

KernelCAT achieves intelligent operator optimization for domestic chips.

Google's AI Advantage: Why Spider Separation is the Only Way to Achieve a Fair Internet

Cloudflare criticizes Google's proposed regulatory rules, urging the separation of Googlebot as an AI crawler.

Google's AI Advantage: Why Spider Separation is the Only Way to Achieve a Fair Internet

Cloudflare criticizes Google's proposed regulatory rules, urging the separation of Googlebot as an AI crawler.

More Disheartening Than the Gold Plunge: Vue 3.6's Build Tools Are Fully Switched to Rolldown 😰

Vue 3.6 unifies all build tools to Rolldown.

More Disheartening Than the Gold Plunge: Vue 3.6's Build Tools Are Fully Switched to Rolldown 😰

Vue 3.6 unifies all build tools to Rolldown.

General

Cubic Joins Vercel Agents Marketplace

Cubic joins Vercel Agent market, offering AI code review.

Cubic Joins Vercel Agents Marketplace

Cubic joins Vercel Agent market, offering AI code review.

Implementing an Interactive 3D Map of China Using Three.js + D3.js (Province Focus / Fly Lines / Light Columns)

Three.js + D3.js is used to implement a 3D interactive map of China.

Implementing an Interactive 3D Map of China Using Three.js + D3.js (Province Focus / Fly Lines / Light Columns)

Three.js + D3.js is used to implement a 3D interactive map of China.

January 31, 2026

Multidomain news, covering international, economic, sports, and technology topics.

January 31, 2026

Multidomain news, covering international, economic, sports, and technology topics.

Short Month, Big Creativity (February 2026 Wallpaper Edition)

Smashing Magazine released free desktop wallpapers for February 2026.

Short Month, Big Creativity (February 2026 Wallpaper Edition)

Smashing Magazine released free desktop wallpapers for February 2026.

Artemis I: 13th Day of Flight

Artemis I mission on its 13th day.

Artemis I: 13th Day of Flight

Artemis I mission on its 13th day.

When Snowflakes Fall in China: A Moment of Transformation for Enterprise AI Strategy ｜ 2025-2026 Data+AI Annual Moments Highlights

InfoQ publishes the 2025-2026 Data+AI Annual Moments Highlights.

When Snowflakes Fall in China: A Moment of Transformation for Enterprise AI Strategy ｜ 2025-2026 Data+AI Annual Moments Highlights

InfoQ publishes the 2025-2026 Data+AI Annual Moments Highlights.

Episode 200: Evolution

Podcast reviews the current state of Japan and the U.S., and discusses AI topics.

Episode 200: Evolution

Podcast reviews the current state of Japan and the U.S., and discusses AI topics.

Primary

OCR Competition Intensifies! Baidu Releases New SOTA OCR Model, Outperforms DeepSeek-OCR2

InfoQ - AI & LLM•Published on January 30

News•Computer Vision•Score: 8/10

OCRPaddleOCR-VL-1.5SOTA modeldocument parsingirregular bounding box localizationBaidu

Source

◆

Memory Prices Surge 170% Year-over-Year, It's Time to Clarify the 'Hidden Costs' in Cloud Bills

InfoQ - Cloud•Published on January 28

News•Cloud Cost Optimization•Score: 8/10

Memory Price IncreaseCloud CostFlexible ComputingHuawei Cloud FlexusResource WasteCPU/Memory Ratio

Source

◆

Comparison between HTTP/2 Based on TCP and HTTP/3 Based on QUIC

ByteByteGo•Published on January 31

Insight•System Design•Score: 7/10

HTTP/3QUICTCPhead-of-line blockingAI predictionprogramming agent

Source

◆

AWS Releases Digital Sovereignty Lens for Well-Architected Framework

AWS - Architecture•Published on January 30

Insight•Cloud Computing Architecture•Score: 7/10

AWSDigital SovereigntyWell-Architected FrameworkComplianceData ResidencyAuditability

Source

◆

Why Is There a Video Tag? Why Do We Still Need a Video Player?

Juejin Frontend Hottest•Published on January 29

Insight•Frontend Development•Score: 7/10

HTML5 videovideo playerHLSDASHVideo.jsfrontend framework

Source

◆

Efficient JSON Data Analysis Using Amazon Redshift SUPER

AWS - Databases•Published on January 27

Insight•Database•Score: 7/10

Amazon RedshiftSUPER data typeJSON analysissemi-structured dataPartiQLdata warehouse

Source

◆

Vercel Sandbox is Now Fully Available, Supporting Execution of Untrusted Code

Vercel News•Published on January 30

News•Serverless Platform•Score: 7/10

Vercel SandboxAI AgentMicro-Virtual MachineServerlessCode Sandbox

⚠️

Notable

1. Vendor Lock-in Risk: Although Vercel Sandbox is powerful, it binds core compute logic to a specific platform. Migrating business operations would involve significant cost and complexity. This is clearly Vercel's strategy to build an ecosystem moat, but it is not the optimal solution for enterprises pursuing multi-cloud or hybrid cloud strategies. 2. Cost Transparency: While it claims to charge based on active CPU time, the actual billing model for complex AI Agent tasks often proves more intricate than expected, potentially leading to hidden costs, especially in high-concurrency and sudden traffic scenarios. 3. Technical Black Box: As a managed service, the specific configurations, resource limits, and security audit capabilities of the underlying Firecracker micro-virtual machines are black boxes for users. When performance bottlenecks or security incidents occur, diagnostic and intervention capabilities are limited.

📈

Perspectives

In the international market, FaaS platforms like AWS Lambda, Google Cloud Functions, and Azure Functions are the traditional representatives of Serverless computing, but they mainly target human-driven business logic. For AI Agent's transient and isolated compute requirements, platforms like RunPod and Modal Labs offer more specialized GPU/CPU cloud services, while Google Cloud's Cloud Run and AWS Fargate have also made inroads into containerized Serverless solutions. The advantage of Vercel Sandbox lies in its integration with web development and AI Agent ecosystems. However, for scenarios requiring high-performance AI inference or custom hardware acceleration, specialized GPU service providers may still be the better choice. In China, Alibaba Cloud Function Compute and Tencent Cloud Serverless Cloud Function are also developing, but platforms specifically offering micro-virtual machine sandboxes for AI Agents are not yet widespread. The Vercel Sandbox model holds unique market potential in China, especially in the AI application development and deployment field.

Source

◆

Aspire 13.1 Release: Bringing MCP Integration, CLI Enhancements, and Azure Deployment Updates

InfoQ - Cloud•Published on January 29

News•Cloud-native Development Framework•Score: 7/10

.NET AspireCloud-nativeAI-assisted developmentAzure deploymentMCP

⚠️

Notable

1. Aspire remains deeply tied to the .NET ecosystem and Azure cloud platform. For teams using non-.NET technology stacks or adopting multi-cloud/hybrid cloud strategies, its applicability and migration costs are still evident limitations, and it cannot become a generic cloud-native development framework in the short term. 2. The effectiveness of AI agents integrated with MCP largely depends on the intelligence and context understanding capabilities of the AI tool itself. If the AI assistant fails to accurately grasp complex business logic and architectural intent, its assistance may remain limited to basic code inspections. 3. As a product within the Microsoft ecosystem, its 'openness' and 'community-driven' nature lags behind truly neutral cloud-native projects like Kubernetes, Istio, etc., potentially limiting its influence within the broader open-source community.

📈

Perspectives

In the field of cloud-native development frameworks, there are numerous mature and active open-source projects globally, such as Kubernetes (container orchestration), Helm (package management), Dapr (distributed application runtime), etc. Aspire, as a Microsoft-led framework, faces more competition from cloud service provider integrated development environments, such as AWS Copilot, Google Cloud Code, etc. However, Aspire focuses more on enhancing local development experience and deep integration with IDEs. In China, with the growing adoption of cloud-native technologies, major cloud vendors are also promoting their own developer toolchains. The updates in Aspire 13.1, particularly in AI-assisted development and Azure deep integration, make it highly attractive within the .NET developer community. To become a universally applicable framework across languages and cloud platforms, it needs stronger ecosystem openness and community neutrality. For teams pursuing technical diversity and avoiding vendor lock-in, a combination of Dapr with other cloud-native toolchains might be a more flexible alternative.

Source

◆

Limited Budget, Technical Gaps: The Most Challenging AI Users Are Small and Medium Enterprise Owners

InfoQ - AI & LLM•Published on January 30

Insight•AI Agent Platform•Score: 7/10

Huawei Cloud FlexusAI AgentSmall and Medium Enterprise AIRAGEnterprise Knowledge Questioning

⚠️

Notable

1. The promise of 'a demo per day, results visible within a week' is overly aggressive. In the ever-changing business processes of small and medium enterprises, deep customization and continuous optimization of AI agents are often long-term efforts. Short-term results may only represent superficial integration. 2. Although it emphasizes 'out-of-the-box' usability, the data cleaning, labeling, and knowledge base construction for enterprises are themselves significant challenges. The Flexus platform's support in these areas and associated costs remain hidden barriers, potentially requiring additional specialized service investments. 3. Faced with 'the most demanding AI users,' the core issue often lies in the complexity and non-standardization of business logic. Even with numerous pre-configured templates, it's difficult to cover all customized needs, making the platform potentially a useless tool that's too general to be effective.

📈

Perspectives

In the domestic market, major cloud providers such as Alibaba Cloud, Tencent Cloud, and Baidu Smart Cloud are actively deploying enterprise-level AI solutions, especially in RAG and AI agent development platforms. For instance, Alibaba Cloud's Tongyi Lingma, Baidu Smart Cloud's Qianfan Large Model Platform, offer a wide range of AI capabilities and development tools. The differentiation of Flexus AI agent lies in its clear focus on small and medium enterprises and scenario-specific implementation, rather than general large model capability output. Compared to international competitors like Azure AI Studio and Google Cloud AI Platform, Huawei Cloud has an advantage in localization services and data security compliance in China. However, to truly win the small and medium enterprise market, it needs a strong ecosystem of partners and sustained operational support to address the 'last mile' challenges, avoiding becoming another 'looks good but doesn't work' platform.

Source

Essential

Zhiyuan Multimodal Large Model Published in Nature, Establishing Autoregressive as a Unified Path for Generative AI

InfoQ - AI & LLM•Published on January 29

News•Large Models•Score: 9/10

ZhiyuanEmu3MultimodalAutoregressiveNext-token predictionNatureAGI

Source

◆

Google's World Model Project Genie Goes Viral: 'Splitting Wood' Brother and Hamersley Take the Stage, Revealing the 60-Second Limit is Not a Limit, Memory is the Major Constraint

InfoQ - AI & LLM•Published on January 30

Insight•World Models•Score: 9/10

World ModelsProject GenieGenie 3Embodied IntelligenceAI-generated Worlds

⚠️

Notable

1. The '60-second limit' and 'diminished dynamism' are its critical drawbacks. While the team explains it as a trade-off between service costs and user experience, this indicates fundamental limitations in the model's ability to generate long temporal sequences and complex dynamic environments, far from achieving true simulation of the real world. 2. The interactive experience remains monotonous, lacking task goals and audio feedback, and is plagued by input latency and character control issues. This is still far from being a 'playable game,' more like a technical demonstration sandbox than a mature product. 3. There are clear restrictions on well-known IPs, with the inability to directly generate related characters, highlighting constraints in copyright, compliance, and complex semantic understanding and content control. This will severely limit its application in commercial entertainment sectors.

📈

Perspectives

World models are one of the hottest directions in the AI field today, with major international players like OpenAI and Meta heavily investing. Meta's I-JEPA, Google's earlier Imagen Video, and Lumiere have made notable achievements in video generation, but Project Genie's emphasis on real-time interactivity sets it apart. Domestic companies such as Baidu Wenxin, Alibaba Tongyi, and Huawei Pang are rapidly advancing in multimodal generation, but specialized world model products are still not clearly visible. Currently, Genie 3 appears to be in the stage of technical validation and exploration, with a long way to go before large-scale commercial deployment. Its high computational costs and hardware dependency pose significant challenges for domestic implementation. In the short term, it will primarily serve as a research tool and professional content creation aid, far from 'disrupting the gaming industry,' let alone replacing mature game engines like Unity or Unreal Engine.

Source

◆

‘The World Has Suffered from CUDA for Too Long!’ KernelCAT Takes the Lead, Enabling Seamless Adaptation for Domestic Chips

InfoQ - AI & LLM•Published on January 30

Insight•AI Chips and Computing Power•Score: 9/10

CUDAKernelCATdomestic chipsAscendoperator development

Source

◆

Google's AI Advantage: Why Spider Separation is the Only Way to Achieve a Fair Internet

Cloudflare Blog•Published on January 30

Insight•Internet Governance•Score: 9/10

GooglebotAI CrawlerSearch Engine MonopolyCopyright ProtectionCMA

Source

◆

More Disheartening Than the Gold Plunge: Vue 3.6's Build Tools Are Fully Switched to Rolldown 😰

Juejin Frontend Hottest•Published on January 31

Insight•Frontend Build Tools•Score: 8/10

Vue 3.6RolldownRollupesbuildMonorepo

Source

General

Cubic Joins Vercel Agents Marketplace

Vercel News•Published on January 30

News•AI Code Review•Score: 6/10

AI code reviewVercelcubicAI Agentcode quality

⚠️

Notable

1. The false positive rate and contextual understanding depth of AI code review have always been pain points. Even though Cubic claims to be able to learn, its accuracy remains questionable when dealing with complex business logic and architectural patterns. Over-reliance may lead to misleading suggestions or overlooking deeper issues. 2. Although it promotes the idea of "reducing manual workload," in actual engineering practices, critical decisions and architectural-level reviews are difficult for AI to replace. In the end, it may become a basic tool for syntax and standard checks. 3. Potential data privacy risks are significant. Entrusting enterprise core code to third-party AI Agents raises challenges in balancing data desensitization and compliance with security standards.

📈

Perspectives

Internationally, tools like GitHub Copilot Enterprise and Code Llama already offer similar AI-assisted coding features, but they mainly focus on generation and support. Tools dedicated to code review include DeepCode (now part of Snyk Code) and SonarQube. In China, major companies like Huawei and Alibaba are also exploring internally, but their external products are not yet fully mature. Cubic may gain initial traction among specific user groups due to its integration advantage with the Vercel platform. However, for broader enterprise adoption, its core competitiveness lies in how it handles private deployment and customized rules, as well as reducing the false positive rate. In the short term, it is optimistic about adoption rates among small and medium-sized teams, but to truly change the code review paradigm, it needs stronger semantic understanding and contextual correlation capabilities.

Source

◆

Implementing an Interactive 3D Map of China Using Three.js + D3.js (Province Focus / Fly Lines / Light Columns)

Juejin Frontend Hottest•Published on January 26

Insight•Frontend Visualization•Score: 6/10

Three.jsD3.js3D MapData VisualizationGeoJSON

Source

◆

January 31, 2026

Global Vision•Published on January 31

News•International News•Score: 5/10

International NewsAI PolicyWinter OlympicsGold MarketSpace Exploration

Source

◆

Short Month, Big Creativity (February 2026 Wallpaper Edition)

Smashing Magazine•Published on January 31

常规更新型•Visual Design•Score: 1/10

desktop wallpapersdesign inspirationFebruary 2026Smashing Magazine

Source

◆

Artemis I: 13th Day of Flight

NASA APOD•Published on January 31

News•Space Exploration•Score: 1/10

ArtemisOrion spacecraftNASAMoon mission

Source

◆

When Snowflakes Fall in China: A Moment of Transformation for Enterprise AI Strategy ｜ 2025-2026 Data+AI Annual Moments Highlights

InfoQ - Big Data•Published on January 28

News•AI Strategy•Score: 1/10

AI StrategyData StrategyEnterprise AIIndustry Trends2026

Source

◆

Episode 200: Evolution

Post-Internet Talk•Published on January 31

News•Podcast•Score: 1/10

PodcastPost-Internet EraAITechnology Commentary