亚洲精品少妇久久久久

Yitu TECH. Raises Over 2 Billion RMB, Accelerates AI Security Transformation: Exclusive Insights from the Team

(Image source: Photo by Lin Zhijia, editor at TMTPOST)

 

After years of low-key activity, YITU Tech., one of the "AI Four Dragons," has begun to accelerate the commercialization of AI in the security field, riding the new wave of AI large models.

At the 2024 China International Public Safety Expo, YITU unveiled several new and upgraded products, including the enhanced "QuestMind^TM LVLM," the Xiaoming AI Agent, and the newly launched "YITU MindV^TM" partnership business brand.

YITU's President, Alex Duan, stated that over the past decade, YITU has developed world-class algorithms and continuously focused on the security industry. This has supported nearly 10 million video intelligence applications, and the company has accumulated significant advantages in algorithms, data, computing power, and AI architecture. Now, with the arrival of AI 2.0, large models have driven the marginal cost of long-tail algorithms close to zero, accelerating the commercialization of AI technologies.

He explained, "By offering highly cost-effective products and empowering partners with the advanced concepts, ideas, and tools of large models, we help our partners transform into providers and operators of scenario-based large model solutions." Duan emphasized that YITU's "MindV^TM" partner business is built on the integration of YITU’s training-push system, Agent architecture, and large model products, which help customers close the loop between data, applications, and services, enabling the development and continuous operation of intelligent applications in various scenarios.

During the event, an exclusive and in-depth conversation took place between XU Yan, Vice President of R&D at YITU Technology, and TMTPOST.

Xu Yan told TMTPOST that there has always been a hope that AI could empower various industries, and that AI large models have solved past challenges, such as high industry costs, high barriers to entry, and the difficulty of generalizing across different scenarios. YITU’s AI technology now matches customer demand scenarios, and its To B/G government and enterprise businesses no longer focus on "project-based" work. Instead, YITU offers integrated hardware and software product solutions, aiming for a high gross margin of over 60%.

According to TMTPOST, after several rounds of business adjustments, YITU has now entered a positive commercial cycle for two consecutive years.

YITU Re-focuses on AI Security Commercialization After Four Waves of AI

Since the term "artificial intelligence" was coined at the 1956 Dartmouth Conference, the world has experienced four waves of AI, with rapid advances in computing technology. In the first two waves, the long-awaited "AI effect" failed to materialize as expected. Whether it was the inability of chip "Moore's Law" to keep up with the computing power demands of AI, insufficient data, weak algorithms, or underwhelming commercialization results, the industry remained skeptical about AI for a long time.

Now, we are experiencing the third AI wave. This time, the "three pillars" of AI—algorithms, computing power, and data—have all broken through simultaneously. From 1997, when IBM’s "Deep Blue" defeated chess champion Garry Kasparov, to 2016, when Google DeepMind’s "AlphaGo" defeated South Korean Go champion Lee Sedol, the last 20 years have seen the real-world application of machine learning, deep neural networks, natural language processing (NLP), and generative AI.

Research organization Gartner updates its "Hype Cycle" every year, tracking the maturity of various technologies. AI technology has gone from a hype cycle through the valley of disillusionment, and today, generative AI is experiencing explosive growth. It has quickly become one of the most widely deployed AI technologies across departments and organizations, emerging from the so-called "Valley of Death" and entering a new phase of U-shaped "inflated expectations." Gartner predicts that generative AI will continue to be a hot trend globally in the next 2-5 years.

As one of the "AI Four Dragons," YITU has gone through a process similar to the "technology maturity curve."

Founded in 2012, YITU Technology focuses on AI chips and algorithms as its core technologies, dedicated to solving the fundamental problems of machine vision, hearing, understanding, and planning. The company aims to provide high-performance, high-density, and universal computing power to support the development and widespread application of AI, meeting the growing demand for intelligent computing in cloud data centers, edge computing, and the Internet of Things.

Over the past decade, YITU has risen with the AI wave, faced a low point around 2019, and is now experiencing a business transformation and upgrade, returning to positive cash flow. To date, YITU has raised over 2 billion RMB in funding from investors including ZhenFund, Sequoia China, Hillhouse Capital, and Yunfeng Capital.

In the context of capital cooling down and the industry returning to rationality, YITU has long focused on AI technology and products, going through business transformation and upgrading, and truly solidifying its competitive edge. Compared to the other three companies in the "AI Four Dragons" — SenseTime, Megvii, and CloudWalk — YITU chose to focus on "security" to address the business needs of smart enterprises and smart cities, offering integrated hardware and software solutions combining chips and algorithms.

Data shows that between 2017 and the first half of 2020, YITU provided products and solutions to over 800 government and enterprise clients in more than 30 provinces, autonomous regions, and municipalities across China, as well as over 10 countries and regions abroad.

In fact, YITU started exploring generative AI large models based on the Transformer architecture early on.

In 2020, YITU launched the pre-trained language understanding model ConvBERT, which, through a new attention module, achieved the same accuracy as Google's BERT model with only 1/10 (10%) of the training time and 1/6 of the parameters. Compared to OpenAI’s GPT-3, it allowed for faster exploration of language model training and reduced the computational cost during model prediction.

With the global popularity of ChatGPT, in July 2023, YITU officially released its visual-centric QuestMind^TM LVLM, which was quickly deployed across national projects. As of now, the Tianwen large model has been implemented and scaled in over 80 projects.

On the technical and business side, by integrating large models (QuestMind^TM), AI Agents (YITU Xiaoming), and NPU chips (QuestCore/QuestSeek), YITU has successfully closed the loop for applying AI technology in the security field.

 

In terms of large models and applications, the latest YITU QuestMind^TM 4.5 system has achieved a 40% improvement in target accuracy for deployment and a 75% reduction in data volume.

At the same time, the YITU QuestMind^TM can quickly adapt to changes in the environment and demands. Compared to traditional machine learning models, which require 1-3 months of data collection and model training, YITU QuestMind^TM has upgraded its pre-trained model. It can achieve a cold start of new algorithms with very few samples in under 1 minute, complete online annotation training within 1 hour, and quickly deploy the system within a day. With just a few minutes per day spent aligning data and simple clicks for correct or incorrect labeling, the algorithm can reach over 90% accuracy within a few days.

Additionally, the YITU QuestMind^TM supports multi-condition scenario deployment, demonstrating extremely high practical application value in areas like urban management, environmental monitoring, and public safety.

For example, in city road sections with frequent traffic accidents, YITU's "QuestMind^TM" large model empowers existing front-end cameras to solve traffic issues. This includes algorithms for behaviors such as running red lights, speeding, failing to yield to pedestrians, cycling without helmets, illegal U-turns, and running over lane markings. The system can be deployed quickly, and after one week of deployment, the accuracy can increase from an initial 60%-70% to nearly 100%, while significantly reducing construction costs.

On the hardware side, at the end of 2022, YITU began mass production of its QuestCore computational power chip. Prior to this, YITU's QuestCore was China's leading custom SoC for cloud-based deep learning inference, used in cloud and edge servers. Now, YITU has developed a multimodal large-model training and inference integrated machine system, which spans a range of multi-scene products for intelligent video applications from 50 to 1000 channels, and is being sold as an integrated software and hardware solution.

On October 23, YITU Tech. also announced its MindV^TM series products, which use low-cost, high-performance solutions to unlock long-tail algorithms and enable long-cycle retrieval.

Xu Yan stated that the essence of AI is research and learning, and the pre-trained AI large model paradigm has changed the entire industry's production model. Based on deep learning and large model technology, through YITU's innovation, cost-effectiveness, and industry application, AI technology has been successfully commercialized in business operations.

Before joining YITU, Xu Yan worked at Dahua, a leading player in the domestic security industry, where he gained many years of experience in the intelligent security business and sales management.

“In the AI 2.0 era of large models, AI agents help solve the 'last mile' problem of large models. With the foundation and model technology in place, everyone can address the challenge of understanding and implementing user needs,” Xu Yan mentioned. He emphasized that YITU is not replacing clients' existing security equipment, but rather utilizing large models to efficiently empower existing front-end devices, enabling these older devices to acquire AI capabilities. This approach avoids redundant infrastructure, improves the accuracy of recognizing dangerous traffic behaviors, and offers advantages in terms of construction costs, application scenarios, algorithm suitability, and dynamic adjustments.

On the commercialization front, past B2B/G business models have faced challenges such as long payment cycles and bad debts. AI leaders such as SenseTime, Megvii (Face++), and iFlytek have been scaling back investments in security and smart city businesses. However, YITU remains committed to focusing on “security” as a key area of commercialization.

Xu Yan told TMTPOST that YITU currently focuses on providing downstream solution environments for the security industry, offering two sales models: B2B partnerships and direct sales. Unlike large-scale project models, YITU provides AI algorithms and computing power terminal software and hardware products. This "pay-on-delivery" model greatly reduces challenges such as bad debts and long payment cycles in the B2B/G market.

From technology to product and then to commercialization, YITU is fully focused on the security sector. Going forward, the company plans to further develop its partner business, specifically addressing the challenges that small and medium-sized enterprises (SMEs) face in practical applications.

Li Xiongwei, General Manager of YITU’s Partner Business, revealed that the company is already working with 37 independent software vendors (ISVs) across eight industries, including finance, firefighting, healthcare, general industrial parks, energy, construction sites, emergency response, and forest protection. YITU plans to expand this to 20 industries and 60 ISVs, and establish an alliance council to vigorously promote certified distributors.

Huge Market Demand in Domestic Smart Security

The domestic smart security industry still has enormous growth potential.

According to industry statistics from the China Security Association and reports from the China Business Research Institute, in 2023, the total output value of the security industry increased from 366.5 billion yuan in 2013 to 1.01 trillion yuan, a growth of 260%, with a compound annual growth rate (CAGR) of 7.13%. However, the penetration rate of smart security in China has only increased from less than 1% in 2015 to around 6.5% in 2022, a growth of just 5 percentage points, far behind the rapid growth of AI technology.

 

This means that the intelligent security sector still holds enormous potential and market demand.

Alex Duan mentioned that since 2016, more and more companies have been discussing AI technology, but the fundamental contradiction in AI implementation has always been the "low production efficiency of long-tail algorithms." Previously, AI had not fundamentally solved this issue, but now generative AI technology has achieved generalization and universality, significantly improving production efficiency.

A supplier of YITU Technology confessed to TMTPOST that in recent years, YITU has shifted from contraction to focus, making many changes. Now, the YITU QuestMind^TM hardware not only works for security scenarios but can also meet customer needs in other sectors such as finance and industry. The pricing is even more advantageous, offering faster and more practical solutions compared to those from some large companies.

While YITU was sharing its business strategy at a partner summit, there were also reports about SenseTime, another of the "AI Four Dragons," making adjustments to its business. SenseTime is focusing on retaining its large model and big device-related businesses while making cuts in its other business lines, including security, autonomous driving, and healthcare.

Clearly, from the "AI Four Dragons" to the "Big Model Six Tigers" (Dark Side of the Moon, Zhipu, Baichuan, Zero One Universe, MiniMax, Jumpspace), the AI industry is undergoing a paradigm shift under the generative AI boom. Whether businesses will become profitable and focus will depend on the survival of AI companies.

Xu Yan, speaking to TMTPOST, emphasized that a high gross margin of 60% is an important indicator for YITU, and this margin comes from market competition. In the security industry chain, hardware takes 40% of the share, integrators take about 10%, AI gets 10%, and cloud computing takes 15%-20%. Therefore, YITU's positioning is in the "backend," providing AI solutions, not acting as an integrator.

When asked whether YITU views companies like Hikvision or Dahua as competitors, Xu Yan denied this. He believes that YITU's position is more about combining software and hardware, without clashing with upstream or downstream companies in the industry chain. YITU leverages AI algorithms and chips as a technology provider, similar to Nvidia's role.

"For YITU, we are very clear about our position in the ecosystem—software and hardware integration. In China, neither software nor hardware is highly valued alone. Only when combined can we achieve both high margins and solve customer problems. That's our entry point. We are not competing with Hikvision or Dahua. We have a good relationship with Huawei. Essentially, YITU uses its own chips and algorithms to solve customers' business problems," Xu Yan told TMTPOST. AI is YITU's biggest market competitive edge.

Xu Yan acknowledged that competition in the security market is becoming increasingly intense. With the economic downturn this year, everyone is focusing more on the essence of commercialization. Previously, some companies used internet "burning money" strategies for their B2B business, but this model doesn't work anymore. YITU not only understands technological innovation but also recognizes the importance of business services. Over the past 10+ years, YITU has deepened its understanding of commercialization.

YITU emphasizes that it has significant advantages in algorithms, data, computing power, and AI architecture. However, in specific scenarios across various industries, its partners have more advantages in domain knowledge and operational services. Moving forward, YITU hopes to accelerate the digital transformation of the security and smart city industries through collaboration with more industry partners to jointly deliver large model solutions for different scenarios.

(Original article first published on TMTPOST, author: Lin Zhijia, editor: Hu Runfeng)

You can copy the link to share to others: //dribblessportsbar.com/node/952