In the AI-Native Era, How Does Baidu's ERNIE Model Ignite Application Innovation?

[baoshi.rao]

On November 21, Baidu officially released its financial report for the third quarter of 2023. The data shows that Baidu's revenue for the quarter reached 34.447 billion yuan, with net profit attributable to Baidu (non-GAAP) reaching 7.3 billion yuan, a year-on-year increase of 23%. Both revenue and profit exceeded market expectations.

Baidu's founder, chairman, and CEO Robin Li stated: "Amid the emerging opportunities in generative AI and foundational models, Baidu continues to strengthen its technology and improve its products, particularly with the launch of ERNIE 4.0, Baidu's most powerful foundational model. We have also leveraged the capabilities of the ERNIE model and ERNIE Bot to revamp consumer and enterprise products as well as our own operations, delivering AI-native experiences while continuously enhancing efficiency."

As Robin Li emphasized: "Baidu remains committed to its AI-centric business and product strategy, laying the foundation for sustained revenue and profit expansion for the ERNIE model and ERNIE Bot ecosystem in the years ahead." Empowered by the ERNIE model, Baidu saw significant growth in AI-native application metrics in Q3: API calls surged exponentially, while user engagement and activity for Baidu's AI-native applications rose rapidly, driving increases in user numbers, online duration, and payment rates.

The rise of large models has placed Baidu, long focused on AI technology, at the center of the era. As a leader in China's AI sector, how Baidu navigates the development of large models and sets an example for the industry is not only crucial for its own growth but also influences the overall direction of China's AI industry.

This year, with the wave of generative AI, Baidu's ERNIE model iteration has been accelerated.

Since the initial ERNIE 1.0 in March 2019, after four years of technical refinement and iterative development, the ERNIE model has now evolved to version 4.0—the most powerful iteration yet. It represents a comprehensive upgrade of the foundational model, with notable improvements in comprehension, generation, logic, and memory capabilities. The model now features a three-tier technical system (foundational-task-industry) and is distinguished by two key strengths: knowledge enhancement and industrial applicability.

Over the past 8 months, utilizing the PaddlePaddle deep learning platform and a 10,000-card computing cluster for training, the model has achieved integrated learning from massive data and knowledge through multi-stage alignment and recyclable training techniques, enhancing its generalization capability and knowledge representation. Since the launch of ERNIE Bot in March, the training algorithm efficiency of ERNIE 4.0 has improved by 3.6 times, with weekly training effectiveness exceeding 98%, and inference performance increasing by 50-fold.

With pioneering technological investments, the 'technology flywheel' of the ERNIE model has begun to show results, demonstrating excellent performance across multiple evaluations and application scenarios. It recently won the 'World Internet Conference Leading Technology Award' at the concluded World Internet Conference.

If the numerous leaderboards in the large model industry are seen as academic 'report cards,' then Baidu's ERNIE model undoubtedly stands out as a top performer. However, technological breakthroughs are just the beginning; the key lies in translating these capabilities into practical value and industrial applications.

Since the emergence of ChatGPT in late 2022, the number of domestic large models has grown rapidly, reaching an astonishing 238, with nearly every model claiming to 'outperform GPT and surpass Llama2' on leaderboards. Yet, alongside this enthusiasm for 'leaderboard dominance,' a critical question remains: What practical value can these large models bring to industries? Regrettably, many models remain stuck at the 'theoretical discussion' stage.

For large models, excessive focus on leaderboard performance may lead to disconnection from real-world applications, turning them into impractical 'exam-takers' that produce 'serious but nonsensical' hallucinations in practice. According to the HalluQA dataset for large model hallucination evaluation, jointly developed by Fudan University and Shanghai AI Lab, only 6 out of 24 mainstream large models achieved a hallucination-free rate above 50%, accounting for just one-quarter of all tested models.

In the evaluation, Baidu's ERNIE Bot topped the list with a 69.33% hallucination-free rate. A higher hallucination-free rate indicates that the ERNIE large model can provide reliability in industrial applications, removing barriers to its widespread industrial adoption. Beyond ERNIE's robust foundational model capabilities and retrieval augmentation, its accumulated experience in real-world industry scenarios has been key to its standout performance.

Clearly, Baidu had set its sights on industrial applications from the very beginning of developing the ERNIE large model. Thus, as the era of large models arrived, Baidu leveraged its first-mover advantage to pioneer the 'ecosystem' battlefield while other companies were still focusing on model development, seeking the 'scenario flywheel' for large models.

Baidu has also been at the forefront of exploring large model applications. Since its public release on August 31, ERNIE Bot has rapidly gained 70 million users, covering 4,300 scenarios and 2,492 applications. Its largest user base consists of young and middle-aged workers in first-tier and super-first-tier cities, with the IT/internet and education sectors being the top industries. The three most frequent use cases are knowledge Q&A, text generation, and coding. ERNIE Bot's smooth interaction and extensive knowledge base make it a powerful assistant for solving work and life challenges.

At Baidu World 2023, Baidu announced a strategic partnership with the Chinese Swimming Association, making ERNIE Bot the 'AI Partner of the Chinese National Diving Team.' It provides an AI-assisted training system to help divers refine every move and perfect each dive. Additionally, the Ministry of Culture and Tourism recognized Baidu's ERNIE large model for innovative cultural product production as one of the top ten digital innovation cases in 2023—the only large model application selected—showcasing its contributions to cultural innovation.

The emergence of large models will lead the industry into the fourth revolution marked by artificial intelligence. Currently, we are in the midst of a transformation driven by generative AI and foundational models, which will completely restructure every industry. It can be said that large models will become a resource every enterprise must use in the AI era. It is already an industry consensus that businesses need to leverage large models to restructure their operations and enhance productivity.

However, the economic and technical costs of developing large models are prohibitively high. "Reinventing the wheel" not only results in significant waste of social resources but also makes it difficult for most general-purpose large models to achieve expected outcomes due to high technical and cost barriers.

Addressing this, Robin Li was the first to point out: "In the AI-native era, we need a million AI-native applications, but not a hundred large models." Only by building a thriving AI ecosystem based on mature large models can we drive a new round of economic growth and unlock immense commercial value.

AI-native applications refer to applications with inherently secure and trustworthy AI capabilities in their design, development, deployment, operation, and maintenance processes, where AI is a natural part of the functionality. Large models and AI-native applications complement each other like operating systems and apps in the mobile internet era. However, the efficiency revolution brought by AI will lead to a major reshuffle in the application market.

To set an example for the industry and encourage more enterprises to participate in building an AI-native application ecosystem, Robin Li explicitly stated: "Baidu aims to be the first company to reconstruct all its products. We will adopt an AI mindset to create AI-native applications." This strategic decision paves a solid path for the industrialization of large models.

Compared to other companies still focusing on large model development, Baidu took the lead by introducing a family of "AI-native applications" based on the ERNIE 4.0 model at last month's Baidu World Conference, including ERNIE Bot, New Search, and New Wenku. These applications have not only achieved comprehensive growth in user numbers and usage duration but also demonstrated new commercialization pathways. For example, New Search has significantly increased Baidu's revenue through "ultimate satisfaction" to enhance user satisfaction and stickiness, "recommendation stimulation" to spark new demands, and "multi-round interaction" for commercial monetization. Ruiliu, Baidu's next-generation intelligent work platform, centers on knowledge management to build an AI-era workflow pipeline, supporting enterprises in improving efficiency and innovation. The Ruiliu Super Assistant enables intelligent office scenarios through natural language interaction.

With successful cases serving as proof, enterprises are increasingly inclined to develop their own AI-native applications using the ERNIE model. Currently, the ERNIE model's API calls exceed the combined total of over 200 other providers, with plugins and API data also surging.

As the era of AI-native applications arrives, large models have become the core engine driving industrial intelligent upgrades. In sectors like e-commerce, healthcare, and finance, large models help enterprises achieve more efficient and precise decision-making and services, undoubtedly bringing significant commercial value.

It is foreseeable that against the backdrop of industries striving for "cost reduction and efficiency improvement," no one wants to be left behind in the intelligent revolution. The demand for AI-native applications will explode.

Addressing five major needs for enterprise large-model implementation, Baidu Intelligent Cloud Qianfan adopts the concept of a "large-model super factory," providing customers with highly efficient and cost-effective heterogeneous computing services, mainstream domestic and international models, and high-quality datasets. Based on the "Qianfan AI-Native Application Development Workbench," it encapsulates common patterns, tools, and processes for developing large-model applications into a single workbench, offering developers a convenient development environment. To meet enterprises' needs for mature AI-native applications, Baidu Intelligent Cloud launched China's first AI-native application store—Baidu Intelligent Cloud Qianfan AI-Native Application Store—providing enterprise clients with a one-stop transaction pathway, significantly improving application selection and procurement efficiency. Currently, the Baidu Intelligent Cloud Qianfan large-model platform has served over 20,000 enterprises in developing industrial models and solutions.

Against the backdrop of strong demand for enterprise AI-native applications, AI-native capabilities may establish new 'rules of the game' in the cloud service sector. Leveraging the deep integration of 'cloud-intelligence synergy' ecosystems, this not only brings new customer leads to Baidu and drives continuous growth and innovation in its cloud business but also creates more commercial opportunities for partners, penetrating industries to achieve ecosystem prosperity and AI democratization.

Final Thoughts

The shift from capability to ecosystem competition is not a transition from the 'first half' to the 'second half' of the large model race but more like an endless marathon. In this competition, the capabilities of large models are merely the starting point, while ecosystem development is the key to sustaining leadership in this long race.

Baidu has not stopped at the 'technology flywheel' but continuously gathers feedback from clients and users to seek the 'scenario flywheel' for further realizing the value of technology. Whether in the development of AI-native applications or the research of foundational large models, Baidu has demonstrated top-tier industry strength and standards.

Large Model Home believes that as more large models shift their focus to industrial value, the future will undoubtedly see more creative and valuable AI-native applications emerge, driving further development and enrichment of AI-native applications, ushering in an era of more open, inclusive, and innovative AI application ecosystems.