To the top net

Value migration driven by AI, redefining the "niche" of digital China's cooperation with Alibaba Cloud

AI is born, and the flying book is really "running through"

Oracle: Data fusion is the key to playing AI

Today, the concept of "data" is far beyond the past, and the concept of "database" is different from the past. AI makes multimodal data fusion more important, which is why Oracle continues to emphasize the importance of integrated databases in the AI era.
More than "fast" and "N-0.5 generation", Sharp Dragon 7 8700G has better solution for opening commercial office

Two heavyweight models, the Reelung 7 8700G and the Core i7-14700, are under the spotlight for a comprehensive and rigorous performance contest.
Schneider Electric and Qianxin jointly create a "technology localization and innovation" circle of friends

Recently, Schneider Electric and Qianxin Group jointly announced the establishment of the Schneider Electric (China) Technology Localization Innovation Center in Yizhuang, adding new momentum to the regional digital economy and information technology application innovation ecological construction.

After analyzing 1000 papers, the Oxford University team found that AI's thinking process was not credible

AI finally learned to remember: the breakthrough of the team of Nanyang Technological University in making the virtual world unforgettable

The research team of Nanyang Technological University has developed the WorldMem framework, which enables AI to have real long-term memory capability for the first time and solves the consistency problem in virtual world simulation. The system stores historical scenes by memory bank and uses intelligent retrieval mechanism to enable AI to accurately reproduce previous scenes and events, even if the interval is very long. Experiments show that it performs well in Minecraft and real scenes, bringing broad application prospects for games, automatic driving, robots and other fields.

AWS strengthens infrastructure strategy and comprehensively upgrades SageMaker to cope with AI competition

artificial intelligence

machine learning

Platform upgrade

2025-07-11

AWS strengthens infrastructure strategy and comprehensively upgrades SageMaker to cope with AI competition

AWS expands its market position by upgrading the SageMaker machine learning platform, adding observation capabilities, connected coding environments, and GPU cluster performance management functions. In the face of fierce competition from Google and Microsoft, AWS focuses on providing AI infrastructure support for enterprises. New features of SageMaker include in-depth insight into the causes of model performance degradation, providing developers with more control over computing resources, and supporting local IDE connection deployment. These updates are mainly derived from customer requirements and are intended to solve practical problems in AI model development.

MTS AI launched: an "intelligent programming assistant" that lets AI write code like a writer writes a novel“

artificial intelligence

Reinforcement learning

code generation

2025-07-11

MTS AI launched: an "intelligent programming assistant" that lets AI write code like a writer writes a novel“

The MTS AI research team proposed the RewardRanker system, which significantly improves the quality of AI code generation through reordering models and iterative self training. This method makes the 13.4B parameter model surpass the 33B big model, performs well in many programming languages, and even surpasses GPT-4 in C++. By introducing difficult negative samples and PPO optimization, the system can select the optimal scheme from multiple code candidates, which lays the foundation for the practicality of AI programming assistant.

The wave rider has no boundaries: Zhiling will expand the blue ocean together in the future

Data compliance and security of enterprises at sea

Those who control waves have no boundaries: crossing mountains and seas to control the distance with wisdom

Those who control waves have no boundaries: change and rebirth to control change

Shengteng AI Developer Summit

Kunpeng Developer Summit

The wave rider has no boundaries: break the boundaries, reconstruct and create new coordinates

Kunpeng Yiteng Developer Conference 2025

Calculate the present and see the future

Implementation of AI Agent manufacturing industry - full analysis of application scenarios from technical scenarios to benefit improvement

Use the sea to find exports, and use technology to navigate the new global track

Use the sea to find exports to control the new variable of "intelligent manufacturing and upgrading"

Advantech x Jetson, accelerating the breakthrough of edge AI industrial application boundary

Use the sea to find exports - intelligent manufacturing upgrading, global layout

Find out the transition from "local manufacturing" to "world smart chain" by going out to sea

How can the telecom industry use generative AI to boost industry innovation?

Data science "tea room - GTC navigation

2024 Amazon Cloud Technology re: Invent

Unlock the overseas growth password by going to sea to find an exit

IBM creates a sustainable future

Break AI "Computing Network" Bottleneck and Accelerate Enterprise Innovation Transformation

Digital navigation

2023 NVIDIA Startups Exhibition Registration Opened

How does BigQuery integrate data and AI to achieve business transformation

artificial intelligence

data processing

Intelligent platform

2025-07-11

How does BigQuery integrate data and AI to achieve business transformation

AI has the potential to transform enterprise insight, but success depends on data quality. Most AI project failures are due to data chaos and dispersion rather than algorithm limitations. Google's BigQuery cloud data AI platform breaks data silos, simplifies governance, and accelerates enterprise AI applications. Through AI automated data processing, real-time analysis is realized, and it is deeply integrated with Vertex AI, so that enterprises can efficiently process structured and unstructured data, and transform intelligent business from vision to reality.

Multimodal large model

Instruction compliance

Training data generation

2025-07-11

The "command following" capability breakthrough of multimodal large model: the team of Shanghai AI laboratory enables AI to understand the requirements of visual tasks accurately like human beings

Shanghai AI Lab, in conjunction with several institutions, launched the MM-IFEngine system, which specifically solves the problem of multimodal AI's "command compliance". The system can automatically generate complex picture instruction training data, and create MM IFEval evaluation benchmark with 400 questions and 32 constraint types. The experiment shows that the trained AI model improves the ability to follow instructions by more than 10%, and significantly improves the ability to understand and execute complex requirements while maintaining the original ability.

Armour UFS 4.1 flash memory promises to improve AI application performance

artificial intelligence

Embedded storage

performance optimization

2025-07-11

Armour UFS 4.1 flash memory promises to improve AI application performance

Armour is testing the latest UFS v4.1 embedded flash memory chip, designed for smart phones and tablets, which can provide faster download speed and smoother device side AI application performance. The chip adopts 218 layer TLC 3D NAND technology, providing 256GB, 512GB and 1TB capacity options. Compared with v4.0 products, the random write performance is improved by about 30%, the random read performance is improved by 35-45%, and the power consumption efficiency is improved by 15-20%. The new standard also adds host initiated defragmentation, enhanced exception handling and other functional features.

The breakthrough research of UCLA: a new era of multi round game of AI dialogue offensive and defensive war

Multiple rounds of dialogue attacks

AI safety protection

Intelligent attack framework

2025-07-11

The breakthrough research of UCLA: a new era of multi round game of AI dialogue offensive and defensive war

This breakthrough research systematically revealed the serious threat of multi round dialogue attacks to AI security for the first time, and developed the X-Teaming intelligent attack framework and XGuard Train protection data set. Research shows that the current AI system has a high miss rate of 98% in the face of carefully designed multi round attacks, but this risk can be significantly reduced through the newly constructed large-scale training dataset, which provides an important tool and new ideas for AI security protection.

artificial intelligence

Automatic programming

Agent mode

2025-07-11

Google Firebase Studio Launches Agent Mode to Realize Automatic Programming

Google released the update of Firebase Studio at the London Cloud Summit, adding Gemini command line interface integration, model context protocol support and "proxy mode". The agent mode provides three AI collaboration levels: the conversational "inquiry" mode is used for brainstorming, the human-computer collaboration agent requires developers to confirm code changes, and the almost autonomous agent mode. Although Google claims that millions of applications have used the platform, it still needs to carefully design prompt words, and non engineer users cannot directly create mature applications.

When AI has permanent memory: MemOS created by the team of Shanghai Jiaotong University makes the big model say goodbye to "amnesia"“

artificial intelligence

Memory management

Large language model

2025-07-11

When AI has permanent memory: MemOS created by the team of Shanghai Jiaotong University makes the big model say goodbye to "amnesia"“

The team of Shanghai Jiaotong University has developed the MemOS memory operating system, which enables AI to have a real long-term memory capability. The system manages three types of parameter memory, activation memory and clear text memory in a unified way, and realizes memory life cycle management and cross type conversion through MemCube intelligent unit. In the LOCOMO benchmark, MemOS achieved the best results in all reasoning tasks, especially in multi hop reasoning and time reasoning.

Google adds image to video generation function for Veo 3

artificial intelligence

image recognition

New algorithm

2025-07-11

Google adds image to video generation function for Veo 3

Google announced on Thursday that it will add image generation video function to its Veo 3 AI video generator through Gemini application. This function was previously provided in the AI video tool Flow launched at the I/O developer conference in May. At present, the Veo 3 video generation function has been launched in more than 150 countries, and is only available to Google AI Ultra and Pro users, with a daily limit of 3 videos. Users can upload photos and add audio descriptions to generate videos. Since its release 7 weeks ago, users have created more than 40 million videos, all with visible and invisible digital watermarks.

4D reconstruction

Asynchronous shooting

Video diffusion model

2025-07-11

A major breakthrough of Shanghai AI Laboratory: using ordinary cameras can shoot ultra-high speed slow lenses, and 4D reconstruction technology subverts traditional photography

The team of Shanghai AI Laboratory proposed an innovative asynchronous shooting scheme, which can achieve high-speed 4D reconstruction with only ordinary cameras. This method improves the effective frame rate from 25FPS to 100-200FPS by staggering the camera start time, and combines the video diffusion model to repair the reconstruction artifacts caused by sparse view angles. The experimental results show that the new method is significantly better than the existing technology in dealing with fast moving scenes, which opens up a new path for low-cost and high-quality 4D content creation.

Nvidia plans to launch AI chips specially designed for the Chinese market

artificial intelligence

chip design

Market strategy

2025-07-11

Nvidia plans to launch AI chips specially designed for the Chinese market

It is reported that Nvidia plans to launch AI chips specially designed for the Chinese market as early as September to cope with the US export restrictions. This chip is based on the Blackwell RTX Pro 6000 processor, which has been modified to comply with the existing AI chip restrictions, and does not include advanced functions such as high bandwidth memory and NVLink high-speed communication interface. Previously, the CEO of Nvidia said that the Chinese market would no longer be included in the revenue forecast, but this change may be short-lived.

Shanghai AI Laboratory and Fudan University: Let AI learn to compare and evaluate like human beings, and redefine the training method of reward model

artificial intelligence

Reward modeling

Reinforcement learning

2025-07-11

Shanghai AI Laboratory and Fudan University: Let AI learn to compare and evaluate like human beings, and redefine the training method of reward model

Shanghai AI Laboratory, together with Fudan University, proposed POLAR method, which is a revolutionary reward model training technology. By letting AI learn to recognize the differences between different strategies rather than memorizing the scoring criteria, POLAR has achieved significant improvement in multiple tasks. The 7B parameter model has exceeded the existing strongest baseline of 72B, providing a new solution to the AI alignment problem.

YouTube's new policy aims to reduce the low quality video content generated by AI

artificial intelligence

natural language processing

Content supervision

2025-07-11

YouTube's new policy aims to reduce the low quality video content generated by AI

YouTube plans to update the monetization policy on July 15 to crack down on "unreal" content, with the main goal of slowing down the proliferation of AI generated content on the platform. The new policy will better identify mass production and repetitive content, including narration story channels with only superficial differences and slide videos with the same explanation. Although YouTube emphasizes that this is only a small update of the existing "duplicate content" guidelines, the industry believes that this move is mainly aimed at the problem of low value AI junk content.

Fudan University Launches BMMR Dataset: Let AI Overcome the Knowledge Gap and Move from Mathematical Genius to Doctor of General Education

artificial intelligence

Multimodal reasoning

Interdisciplinary assessment

2025-07-11

Fudan University Launches BMMR Dataset: Let AI Overcome the Knowledge Gap and Move from Mathematical Genius to Doctor of General Education

The team of Fudan University released the BMMR dataset, which is the first large-scale interdisciplinary multimodal reasoning evaluation benchmark, including 110000 university level questions covering 300 disciplines. The research finds that even the most advanced AI models have limited performance in interdisciplinary reasoning, which reveals the important blind spot of current AI development. This data set supports both Chinese and English, and is equipped with a special reasoning process evaluation tool, which provides important support for AI to develop into general intelligence.

Citrix returns to the mainstream virtualization market, but acknowledges that the product is not ready

cloud computing

Virtualization technology

Market repositioning

2025-07-11

Citrix returns to the mainstream virtualization market, but acknowledges that the product is not ready

Citrix announced its return to the mainstream virtualization market through XenServer products, although the company admitted that the products were not fully ready. Cloud Software Group said XenServer is expanding its support to cover various workloads. Citrix basically abandoned XenServer as a mainstream virtualization product in the early 2010's. The senior director of product management believes that the current virtualization market is undergoing unprecedented changes. In particular, radical licensing changes of major manufacturers have put pressure on IT budgets, providing Citrix with an opportunity to return to the market.

Robots can also "predict the future"? The team of Shanghai Jiaotong University let AI imagine before acting, and the accuracy rate soared to 76.7%

Visual Language Action Model

World knowledge prediction

Diffusion converter

2025-07-11

Robots can also "predict the future"? The team of Shanghai Jiaotong University let AI imagine before acting, and the accuracy rate soared to 76.7%

The DreamVLA system jointly developed by Shanghai Jiaotong University and other colleges and universities enables robots to "predict the future", guide actions by predicting dynamic changes, spatial structures and object semantics, achieve a success rate of 76.7% in real world tests, and set a new record of 4.44 average task length in the CALVIN benchmark of the simulated environment. The system uses a block attention mechanism to coordinate the three predictive capabilities, and uses a diffusion converter to generate accurate actions from noise, representing an important breakthrough from "reactive" to "predictive" robot intelligence.

Docker launches new functions to support AI agent development

artificial intelligence

Containerization technology

Cloud deployment

2025-07-11

Docker launches new functions to support AI agent development

Docker released major new functions aimed at simplifying the construction, operation and deployment of agent AI applications. The company expanded the Docker Compose tool to support large-scale deployment of AI agents and models, and launched Docker Offload service, allowing developers to transfer AI workloads to the cloud. The new function also supports the secure connection of the model context protocol gateway, and integrates with Google Cloud, Microsoft Azure and other partners.

"Nutritionist" of AI training data: How does the Chinese Academy of Sciences make machine learning eat better

data processing

Programmatic editing

AI training optimization

2025-07-11

"Nutritionist" of AI training data: How does the Chinese Academy of Sciences make machine learning eat better

This REFINEX framework, developed by the team of the Institute of Computing, Chinese Academy of Sciences, solves the problem of low quality of AI training data through innovative deletion editing methods. This method first uses expert models to generate high-quality text, and then extracts and deletes operations to train small editing models, which not only ensures data quality but also improves processing efficiency. The experiment shows that REFINEX improves the performance of 750M parameter model by 2.6% - 7.2%, and significantly reduces the training cost.

The new EU AI rules are the terms that are hated by technology giants

artificial intelligence

transparency

New rules and regulations

2025-07-11

The new EU AI rules are the terms that are hated by technology giants

The EU issued detailed rules for the implementation of the AI Act, requiring Google, Meta, OpenAI and other companies to promise not to use pirated materials to train AI, and to disclose in detail the source of training data and the concept of model design. The new regulations also require companies to respect the payment wall and website crawler restrictions, disclose the total energy consumption of training and reasoning, and report security incidents to the EU AI Office within 5-10 days. Illegal enterprises may face a fine of 7% or 3% of annual sales.

ArtifactsBench: Tencent's hybrid team redefines the evaluation criteria for AI code generation, allowing machines to write interface programs that can be seen and used

artificial intelligence

Visual interaction code generation

Automatic evaluation system

2025-07-11

ArtifactsBench: Tencent's hybrid team redefines the evaluation criteria for AI code generation, allowing machines to write interface programs that can be seen and used

Tencent's hybrid team launched ArtifactsBench, the world's first comprehensive benchmark system dedicated to evaluating the quality of AI generated visual interaction code. The system includes 1825 real application scenario test tasks, which are innovatively evaluated from ten dimensions, such as functionality, aesthetics, and user experience, through actual running code, dynamic screenshots, multimodal AI judges, etc. The experimental results show that the consistency with the judgment of human experts is more than 90%, and the consistency with the industry's gold standard WebDev Arena is 94.4%, setting a new standard for AI code generation capability evaluation.

Recommended information

Mail subscription

AI Apocalypse

Value migration driven by AI, redefining the "niche" of digital China's cooperation with Alibaba Cloud

AI is born, and the flying book is really "running through"

The wave rider has no boundaries: Zhiling will expand the blue ocean together in the future

Data compliance and security of enterprises at sea

Those who control waves have no boundaries: crossing mountains and seas to control the distance with wisdom

Those who control waves have no boundaries: change and rebirth to control change

Shengteng AI Developer Summit

Kunpeng Developer Summit

The wave rider has no boundaries: break the boundaries, reconstruct and create new coordinates

Kunpeng Yiteng Developer Conference 2025

Calculate the present and see the future

Implementation of AI Agent manufacturing industry - full analysis of application scenarios from technical scenarios to benefit improvement

Use the sea to find exports, and use technology to navigate the new global track

Use the sea to find exports to control the new variable of "intelligent manufacturing and upgrading"

Advantech x Jetson, accelerating the breakthrough of edge AI industrial application boundary

Use the sea to find exports - intelligent manufacturing upgrading, global layout

Find out the transition from "local manufacturing" to "world smart chain" by going out to sea

How can the telecom industry use generative AI to boost industry innovation?

Data science "tea room - GTC navigation

2024 Amazon Cloud Technology re: Invent

Unlock the overseas growth password by going to sea to find an exit

IBM creates a sustainable future

Break AI "Computing Network" Bottleneck and Accelerate Enterprise Innovation Transformation

Digital navigation

2023 NVIDIA Startups Exhibition Registration Opened

Gartner: Two Strategies to Accelerate the Enterprise's Going to Sea under the Background of Multi culture

Gartner: Use AI agents to enhance business value of enterprises

Gartner: By 2030, guardian agents will occupy 10% - 15% of the agent AI market share

Gartner Releases Five Basic Principles for Enterprises to Build Intelligent Applications

Gartner Releases Five Steps for Enterprises to Build AI Ready Workforce

Gartner: Global GenAI expenditure will reach 644 billion dollars in 2025

Gartner releases 2025 data and analyzes important trends

Gartner: Three measures taken by Chinese enterprises to bridge the gap in cloud skills between infrastructure and operations

If you are eager to know the latest products and technical information in the IT field, subscribing to the technical email on the top network will be one of the best ways for you

AWS re: Inforce: Philadelphia not only strives for goals, but also strives for generative AI security

In Madrid, the gate of Europe, we once again witnessed the "flywheel" that SAP activates enterprise growth

Hu Ning, an expert member of IEEE: dispel the "shadow" of AI illusion and see the value anchor of "super individual"

Generative AI spawns new species, SaaS CRM sets off the second half of the war of change

Follow the official account

Follow the official microblog

Follow the official Himalayas

Friendly Links

Industry hotspot: