In the crisp autumn of October, the 2024 Beijing Security Expo quietly concluded, marking the end of another significant event in the security industry. Although the scale of this year's event was smaller than in previous years, it still carried the hopes and sentiments of the security community. The Expo serves as a window to reflect on the industry's transformations over time, and, at the same time, provides a glimpse into the future trends of the industry in the new era.

On the technology front, security devices are continuing to make breakthroughs in ultra-high definition, intelligence, full-color night vision, low power consumption, and integrated functionalities. On the application front, many security manufacturers are expanding their enterprise-level business, integrating video-based digital solutions into a variety of sectors, providing security management, production management, and operational management solutions to empower businesses' digital transformation.

Looking back, the security industry has evolved significantly since the 1990s, transitioning from analog to digital, then to networked systems, and now to an intelligent era. Early intelligent security applications began with video structuring to enable recognition and detection of people and vehicles. As artificial intelligence (AI) has increasingly integrated into the field, the security industry has entered an AI era, enabling systems to "see" and "understand" a broader range of targets and behaviors.

If previous industry transformations were focused on improving the physical capabilities of monitoring devices—i.e., the "visibility" of the system—then in the AI era, security systems are deepening their ability to "understand" through the fusion of AI with the inherent perception and IoT attributes of security devices. Security systems are shifting from basic protective measures to monitoring scenarios in multiple vertical industries, such as safety production, quality monitoring, and environmental supervision.

This shift indicates that security business, traditionally dominated by government projects, is now pivoting towards the enterprise and commercial markets. Security companies are expanding their product and technological capabilities into new, broad applications, with business reach extending beyond traditional security. The combination of AI systems, large models, and security technology is unlocking new application scenarios.

The Need for Large Models

Over the past 8-9 years, while AI technologies based on deep learning have continuously improved, many industries still have numerous long-tail scenarios that require AI adaptation. This challenge—matching AI to the diverse needs of various vertical industries—is one of the biggest obstacles to AI's adoption in the security industry, and a significant issue facing digital transformation opportunities.

Alex Duan, President of YITU Technology and Deputy Leader of the AI Expert Group at the China Security Association, believes the problem lies in a huge gap between demand and supply. The production efficiency of long-tail algorithms is low, failing to meet the need for smart security.

He explains: “On the demand side, companies want fine-grained management and all-element perception, but the supply side can only provide specific algorithms and attributes. On the demand side, businesses want AI that delivers fast results and quick iterations, but the supply side relies on vast amounts of training data, which is slow to gather. Businesses also want algorithms that can handle complex scenarios, but the algorithms available are not adaptable and suffer from high false alarm rates. Moreover, different scenarios require different rules and customizations, but there are too few algorithm engineers to meet the demand. These are technical bottlenecks, and breaking through them will lead to further development.”

To meet the highly fragmented digital application demands of long-tail markets, large models have become an emerging tool for AI deployment. Their introduction marks a new era for the security industry. At the 2024 Beijing Security Expo, the updates to YITU’s Tianwen 4.5 model and Uniview’s Wutong 2.0 model are signs of the industry approaching a new turning point.

The Arrival of "AI Security 2.0"

YITU Technology, one of the few AI companies at the Expo with large model capabilities, made a big splash with the launch of its "Tianwen Large Model 4.5." Duan Aiguo explained at the press conference that the technological innovation led by large models has dramatically lowered the marginal cost of long-tail algorithm production, bringing it close to zero. This breakthrough marks the arrival of the "AI Security 2.0" era.

Core Changes Brought by the AI 2.0 Revolution: Compared to traditional deep learning methods, large models and multi-modal AI use a self-supervised learning mechanism, demonstrating excellent performance across domains and multi-scenario tasks. The key changes include:

  1. Upgraded AI Production Efficiency: The mechanism of general data pretraining and domain-specific data fine-tuning significantly improves adaptability across scenarios and domains, with new algorithm output efficiency shifting from monthly to daily cycles.

  2. Upgraded Scene Unlocking: AI now perceives, locates, and evaluates in 3D space and 4D time-space, with full scene and full element perception capabilities. This allows for the intelligent management and operation of security in production and operations, beyond basic security measures.

  3. Upgraded Interaction Experience: Multi-modal data unified representation enables cross-modal data validation and interaction, shifting from “tag filtering” to natural language interaction. This increases semantic understanding and video analysis capabilities, allowing security systems to more accurately identify and predict potential threats.

  4. Democratization of AI: With Agent AI models reasoning causal relationships, from quick thinking to slow thinking, AI is becoming more accessible, reducing the barrier to entry and making it easier for everyone to engage with AI algorithms.

The historical significance of large models lies in their departure from supervised learning in deep learning. Large models use a Transformer-based self-supervised learning mechanism that allows general data pretraining, domain-specific fine-tuning, and transfer learning to adapt to multi-scenario tasks. The breakthrough of Transformer-based multi-modal AI lies in its significant improvement in cross-domain generalization and adaptability.

YITU’s QuestMindTM LVLM, with its ability to quickly adapt to environmental and demand changes, outperforms traditional machine learning models that require 1-3 months to collect data and train. The upgraded QuestMindTM model can cold-start new algorithms with minimal samples within 1 minute, complete online annotation training within 1 hour, and quickly deploy models within a day. With just a few minutes of data alignment and simple corrections, the model can reach over 90% accuracy in just a few days.

Alex Duan has emphasized that large models will push the security industry to the forefront of technological application, accelerating the pace of digitalization and ushering the industry into a new phase of rapid development.

Commercial Model Revolution: The Rise of MaaS (Model as a Service)

The introduction of large models has significantly enhanced the generalization and adaptability of AI models across domains. However, the demand for lightweight model deployment has grown exponentially, and the MaaS (Model as a Service) model is beginning to enter the security industry’s vision.

MaaS provides integrated solutions for AI models and related services, bringing a new commercial model and growth opportunity to the security industry. Security companies are shifting from hardware sales to becoming MaaS service providers with an operational focus.

Advantages of the MaaS Model:

  • Flexibility and Scalability: The MaaS model can be flexibly configured and adjusted to meet the needs of various complex scenarios and business demands.
  • Efficiency and Cost-effectiveness: By sharing and reusing large model resources, MaaS lowers research and deployment costs while improving resource utilization.
  • Technical Support and Continuous Improvement: MaaS models are typically supported by professional teams that ensure high availability and continuous optimization of the models.

MaaS is expected to become the optimal solution for scenario-based intelligent solutions, with its flexibility, cost-effectiveness, and efficiency.

For AI + Industry implementation, there are three key requirements:

  1. Continuous innovation and in-depth scenario exploration to enhance the intelligence level of models, meet complex and diverse business needs, and especially apply long-tail algorithms intelligently to address niche market demands.

  2. Optimizing algorithms and improving model efficiency to provide cost-effective intelligent products, ensuring the sustainability of the business logic—this is especially crucial for large-scale deployment and application.

  3. Building a complete business loop and ongoing operations to ensure the stability and reliability of models and algorithms. Establishing and maintaining business credibility is the foundation for long-term cooperation and customer trust.

 

MaaS Model Gives Birth to the "Industry AI System’s Hexagonal Warrior"

Undoubtedly, the paradigm shift in AI technology has led to a transformation in business models, and the MaaS (Model as a Service) model will inevitably depend on collaboration and complementary strengths among various stakeholders.

"A truly practical and complete AI system does not only rely on core elements such as data, algorithms, and computing power, but also requires the comprehensive support of AI architecture, domain expertise, and operational services. The successful deployment of AI within industries hinges on the full integration of these six core capabilities: algorithms, data, computing power, AI architecture, domain expertise, and operational services."

In this process, top-tier AI-native technology companies like Yitu, which have significant advantages in algorithms, data, computing power, and AI architecture, can provide a solid foundation for the application of large models. On the other hand, industry partners possess deep industry knowledge and technical know-how in their respective fields. They have a thorough understanding of industry needs and form complementary advantages with technology providers. As a result, the "last mile" of large model deployment for specific industry scenarios will be increasingly filled by industry partners.

At the launch event, Alex Duan, President of Yitu, introduced the concept of the "Industry AI System’s Hexagonal Warrior." He explained, "Over the years, Yitu has accumulated significant advantages in algorithms, data, computing power, and AI architecture. However, in specific industry scenarios, our partners' advantages in domain knowledge and operational services are even more pronounced. When Yitu collaborates with more industry partners, we can combine our strengths to create a powerful 'Industry AI System’s Hexagonal Warrior,' jointly providing a complete solution for large model deployment in various scenarios and driving the digital transformation across industries."

To help more partners become true "Industry AI System’s Hexagonal Warriors," Yitu officially launched its new partner business brand, "Yitu MindVTM" at the event. The brand is committed to empowering partners through cost-effective products and advanced ideas, concepts, and tools from large models, helping them transform into "providers and operators of scenario-based large model solutions" and win together in the new era of large models.

In the future, with the continuous expansion of application scenarios and the ongoing upgrade of the complexity of digital technologies and their application processes, the "industry AI system's hexagonal warrior," built through the collaboration of AI-native enterprises and industry solution providers, will inevitably become the ecological cooperation consensus in the MaaS (Model as a Service) era. Through close collaboration within the industry ecosystem, a comprehensive and refined large-model solution for various scenarios will be formed, driving the digital intelligence process across industries.

You can copy the link to share to others: //dribblessportsbar.com/node/953