2024: In-Depth Analysis of Large Model Industry Application From "Hundred Model War" to an "Application-First" Era

2024: In-Depth Analysis of Large Model Industry Application From "Hundred Model War" to an "Application-First" Era2024 has seen the development and industrial application of large language models (LLMs) become a global focus. The domestic LLM market has transitioned from an intense "Hundred Model War" to a new phase focused on practical application and deployment

2024: In-Depth Analysis of Large Model Industry Application From "Hundred Model War" to an "Application-First" Era

2024 has seen the development and industrial application of large language models (LLMs) become a global focus. The domestic LLM market has transitioned from an intense "Hundred Model War" to a new phase focused on practical application and deployment. Recently, Wu Tian, Vice President of Baidu Group and Deputy Director of the National Engineering Research Center for Deep Learning Technology and Application, appeared on CCTV-2's "Dialogue" program. She joined Chinese Academy of Sciences academician He Jifeng, Professor Chen Haibo (Chairman of the Technical Guidance Committee of the OpenHarmony project group and specially appointed professor at Shanghai Jiao Tong University), and representatives from other companies to discuss the current state, challenges, and future of LLM development and industrial application.

Wu Tian highlighted the dual nature of LLMs: their capacity for both hallucination and creativity. In scenarios demanding creativity, the powerful generative capabilities of LLMs are a significant advantage. However, when integrating with real-world industry problems, stability and accuracy become paramount for effective problem-solving. Technically, mitigating hallucinations involves two main approaches: optimizing the foundation model to fundamentally reduce the frequency and severity of hallucinations; and incorporating techniques like retrieval augmentation and agents to enhance factual consistency within the application.

According to the China Academy of Information and Communications Technology's (CAICT) "Global Digital Economy White Paper (2024)", there are currently 1328 LLMs globally, with China accounting for a significant 36%. LLM technology is rapidly evolving. Wu Tian shared application data from Baidu's Wenxin LLM, showcasing this rapid growth. Daily calls to Baidu's Wenxin LLM have surged past 1.5 billion, a 7.5-fold increase from 200 million in May and a staggering 30-fold increase compared to 50 million in the same period last year. Furthermore, publicly available data indicates that the number of developers using PaddlePaddle and Wenxin has reached 18.08 million, serving 430,000 companies and creating over 1.01 million models. This rapid growth clearly demonstrates the immense application potential of Wenxin.

Despite the considerable scale of LLM applications, Wu Tian emphasized that true deployment across various scenarios requires further development and refinement. Many applications are currently in a stage of quantitative accumulation. She observed a shift in AI usage from improving efficiency in individual scenarios to progressively optimizing entire business processes. A systemic approach moving from point to line, line to surface, and surface to volume is crucial for achieving qualitative change.

Regarding whether the final hurdle for LLM industrial application is the "last kilometer" or "last hundred kilometers," Wu Tian suggested this depends on the industry's digital foundation and the degree to which its business processes are amenable to digital abstraction. For internet companies, where product applications and AI benefits are closely aligned, deployment might only require half a kilometer. Industries like finance, with a higher level of digitalization, face challenges primarily in system integration and deployment, representing a distance of about one kilometer. However, sectors such as agriculture, with weaker digital foundations, face far more complex challenges, potentially requiring tens or even hundreds of kilometers to achieve large-scale deployment.

Wu Tian further analyzed two major challenges in applying AI across various industries. The first is identifying genuine needs based on real-world scenarios. While numerous AI applications have emerged over the past few years, many fail to address core issues, hindering the realization of AI's full potential. To address this, Baidu established the AICA talent cultivation program six years ago, now in its eighth year. This program aims to cultivate professionals proficient in both business and AI, facilitating better integration of AI with real-world problems.

The second challenge is the lack of specialized data. Data reflecting an understanding of business logic is scarce in naturally occurring data, yet it is crucial for enhancing the ability of models to perform complex tasks in real-world scenarios. Building such datasets requires collaborative efforts from all sectors to support widespread AI adoption.

Computational power remains a significant challenge. Wu Tian acknowledged the exponential increase in computing needs with growing model sizes. Efficient utilization of large-scale computing power requires centralized construction, optimized scheduling, and improved utilization rates. Baidu, through joint optimization of its PaddlePaddle deep learning platform and Wenxin LLM, has achieved a 5.1-fold increase in training efficiency and a 125-fold improvement in inference efficiency, with ongoing optimization efforts. Diversification of computing power is also crucial, especially for large-scale inference and scenario-specific training, reducing costs and enhancing efficiency. Baidu actively promotes hardware standardization, with PaddlePaddle supporting over 60 chip series, ensuring diverse computing power support.

Currently, the domestic LLM market is undergoing a transition from the "Hundred Model War" to a "survival of the fittest" phase, shifting from competition based on parameters to a focus on practical application and deployment. Various industries are adapting LLMs to their specific scenarios, experiences, standards, and data, creating an extensive application ecosystem.

Analysts predict that LLMs will further evolve in 2025, achieving new heights in understanding, generation, and interaction capabilities, accelerating integration across various sectors. By 2025, LLM applications are expected to penetrate numerous vertical industries, forming more mature industry solutions. LLMs are poised to bring about even more profound changes to society, the economy, and technology.

Disclaimer: The content of this article is sourced from the internet. The copyright of the text, images, and other materials belongs to the original author. The platform reprints the materials for the purpose of conveying more information. The content of the article is for reference and learning only, and should not be used for commercial purposes. If it infringes on your legitimate rights and interests, please contact us promptly and we will handle it as soon as possible! We respect copyright and are committed to protecting it. Thank you for sharing.(Email:[email protected])