Produced by | Bullet Finance

Art Editor | Qianqian

Reviewer | Song Wen

This year is known as the "first year of the explosion of large model applications." In 2022, ChatGPT emerged suddenly. In just over two years, the technological wave sparked by large models has swept through various fields and has been hailed as ushering in the "Fourth Industrial Revolution."

Coincidentally, this year marks the 30th anniversary of the development of the Internet in China. At the same time, the development of the Internet has entered its second half. The revolution of Artificial General Intelligence (AGI) centered on large models is essentially another innovation in computer science, which is bound to inject new momentum into the leapfrog development of the Internet industry.

Currently, large models are entering a period of commercial explosion, and application implementation has become a new keyword in the current AI race. New business stories are staged in turn.

On September 25th and 26th, during the 2024 Baidu Cloud Intelligence Conference, special forums on the Internet and embodied intelligence were held. At the forum, executives from more than ten technology companies shared their practical experience in applying large models and discussed the situation of enterprises in the era of Artificial Intelligence Generated Content (AIGC).

1. The influx of large models, and innovators' choices matter for the future

"This is an exciting era," said Wei Zhang, General Manager of Baidu Intelligent Cloud's Pan-Technology Business Department in his speech at the Internet Special Forum, "Every major technological revolution brings epic changes, and so does the large model."

The AI application wave sparked by artificial intelligence is stirring up the entire business world. Innovators on the application side have turned their attention to large models. Leading manufacturers are even more ambitious and eager to compete.

Just as in the early days of the Internet, artificial intelligence also needs to go through a period of wild development with mixed results. During this period, the choices of innovators are related to their long-term development. Among the myriad of large models, choosing a truly advanced large model will allow innovators to run faster and further.

This is also the original intention of Baidu's 2024 Baidu Cloud Intelligence Conference's special forums on the Internet and embodied intelligence.

(Photo / Wei Zhang, General Manager of Baidu Intelligent Cloud's Pan-Technology Business Department)

Wei Zhang observes that amidst the wave of artificial intelligence large models brought about by epic changes, algorithms, data, and computing power continue to iterate, giving rise to many new models and scenarios.

"Technologies and scenarios related to spatial intelligence seemed to require some time last year, but they are now almost upon us," said Wei Zhang. "This year has seen the emergence of many multimodal large model products, including video and 3D generation, leading to a wave of innovative companies. There have also been corresponding breakthroughs in many vertical applications, such as life sciences and new materials. Meanwhile, in various application scenarios such as education, recruitment, and legal services, large models have been explored scenario by scenario. This year, we have also witnessed many tangible implementations that have improved efficiency and restructured our products."

Li Shiyong, Principal Architect at Baidu Intelligent Cloud, found that as large models evolve towards scale, infrastructure still needs to address many issues related to computing power and models.

Innovators are acutely aware that large model training involves large scales and long time periods. If hardware, system configuration, and software failures are superimposed, it can easily "cause training interruptions and prevent clusters from training continuously and effectively."

Another pain point is that AI infrastructure covers a wide range of technical fields and is systematically complex. How can training tasks be performed in a targeted manner to maximize the efficiency of computing power? This requires "combining hardware systems with targeted software optimization for various modalities" to meet the performance requirements of the model side. Baidu's Baihe Heterogeneous Computing Platform can accelerate the training and inference performance of some models by more than 100%.

At this point, some innovators have avoided detours by choosing advanced large models and large model platform toolchains.

2. Amid the wave of the Internet's "intelligent transformation," positive-sum games have become the dominant trend in innovation

The Industrial Internet is seen as the second half of the Internet, which coincides with the arrival of the revolution in Artificial General Intelligence (AGI).

To date, the integrated development of the Internet and artificial intelligence has seen almost all significant innovations and breakthroughs in the AI field take place in the cloud, giving rise to numerous AI technologies that either facilitate breakthroughs in forward-looking fields or change people's lives and production methods.

Against this backdrop, what challenges and opportunities do innovators face?

"We are currently in an era of large model explosion and an era of intelligent exploration across all industries," said Xiaoming Zhang, Vice President of Technology at BioMap.

BioMap is an AI pharmaceutical company founded by Baidu's founder Robin Li. Xiaoming Zhang noted that the well-known pain points in the biopharmaceutical scenario are "a decade-long research and development cycle and a research and development cost of $1 billion." The situation has now worsened, necessitating an accelerated exploration of the AI path.

A typical case is the industry's first $1 billion project based on a biological large model, "demonstrating the determination to invest in this industry."

Xiaoming Zhang revealed that currently, BioMap relies on massive life science big data maps, a cross-modal large model family, and a bio-AI fusion computing engine, built on Baidu Baihe's ultra-large-scale heterogeneous clusters, to create a multimodal life science large model platform to support comprehensive intelligence in industries such as biopharmaceuticals, life health, and scientific research education.

Liu Yuyang, Founder and CEO of Huanliang Technology, mentioned that the biggest challenge in applying AI technology to new material research and development is the issue of data "sparsity," where data acquisition in industrial scenarios in the materials field is costly, time-consuming, and has a low success rate. The limitations of simulation in cross-scale computing and computing power constraints place higher demands on integrating AI into the material research and development process.

Huanliang Technology's launched smart all-in-one machine, combined with Baidu Intelligent Cloud's AI computing power, can autonomously generate materials through AI predictive modeling, improving efficiency in material research and development. Building on this, combined with Baidu Baihe's AI Heterogeneous Platform and Huanliang's self-developed material informatics platform, efficient screening of material formulations that meet target requirements can be performed from millions of combinations.

"In 2020, when I first started working on large models, they weren't yet popular, and it was difficult to explain them to outsiders," recalled Zeng Guoyang, CTO of Mianbi Intelligence. "After ChatGPT emerged, it became much easier to explain."

Mianbi Intelligence is a company focused on large models on the client side and has achieved many breakthroughs in this field, including the previously rumored open-source model "MiniCPM-Llama3-V 2.5," which was allegedly plagiarized by Stanford University in the United States.

Additionally, Mianbi Intelligence cooperates with Accelerated Evolution in embodied intelligence and is also a partner of Baidu Intelligent Cloud's large model end-cloud collaboration solution for client-side models. They have successfully run client-side models on their partners' robotic devices. Going forward, Baidu Intelligent Cloud will work with Mianbi Intelligence to jointly optimize the cloud-client collaboration solution, integrating intent distribution models to synergize client-side and cloud-side large model capabilities, enhancing client-side response efficiency, and helping embodied intelligent robots complete more complex tasks.

DreamTech is an AI startup focused on native 3D generation and launched a 3D generation large model similar to the Sora architecture in May: the world's first native 3D-DiT large model, Direct3D.

Zhang Feihu, CEO of DreamTech, said that the next more important modality is 3D. "3D directly connects the physical world and allows for real-time interaction. Starting from the 3D modality, we can realize a world model and Artificial General Intelligence in the future."","We have extensively used large language models and multimodal models to assist in data processing during 3D generation training," said Zhang Feihu. "Data synthesis is an indispensable solution in the process of building models or developing agents." He believes that this represents an opportunity for startups to overtake their competitors.

At the forum, Baidu Intelligent Cloud once again elaborated on its strategic value proposition of "Cloud Intelligence Integration," expressing its commitment to accelerating the "intelligent transformation" of the Internet industry, walking hand in hand with innovators, and jointly building an intelligent landscape in the AI era.

Positive-sum games in innovation are the key to breaking the deadlock within the industry. Amidst the whispers of technological advancements, the sound of progress echoes loudly.

3. Collaborative scenario applications: How can innovators seek out their "superheroes"?

Amid the wave of the AIGC era, the Internet industry is embracing its best opportunity for leapfrog development.

At this juncture, advanced large models are akin to "superheroes." How can innovators seek out their own "superheroes"?

Large model products are flourishing, but after experiencing them, application-side innovators often encounter many "pits."

For example, some are difficult to adapt to local conditions and find suitable paths. Others fail to improve operational efficiency and establish competitive advantages. Still, others cannot transform from "usable in every industry" to "accessible to everyone." This is partly related to the early stage of large model development and partly to the complexity and variability of downstream applications.

Balancing the ideal and reality of AIGC implementation and traversing the "last mile" of large model application implementation is crucial.

This special forum provided some answers.

"AI has ushered in the era of equal access to knowledge, promoting the large-scale expansion of high-quality educational resources," said Yang Yujian, CTO of Shanghai Juxian Network.

Founded in 2019, Shanghai Juxian Network specializes in ExamBao, with over 60 million users and more than 3 billion test questions to date. "The original cost of manually creating an analysis for a single question averaged 1.5 yuan, but now the cost of AI-generated analysis has been reduced to 0.03 yuan. With machines generating content 24/7, overall content production efficiency has increased by over 1000 times," revealed Yang Yujian.

Yang Yujian believes that users, data, and scenarios must be effectively integrated to form a closed loop. This is attributed to several core business scenarios and processes jointly built by the company and Baidu Intelligent Cloud, such as vectorization, Baidu's search enhancement plugin, SFT, and processes.

Yingteng Education is a company focused on medical examination training, founded in 2005. Currently, it has 16 million structured medical data points and a medical education question bank with 40 million questions, generating 16 billion answer data points.

Having experienced the transition from the Internet to mobile internet and now the AI era, Lantao Lan, Founder of Yingteng Education and Acting President of China High-Tech Group, believes that "using AI to generate test questions will unleash astonishing competitiveness in vocational education. Meanwhile, if enterprises fail to keep up with the pace of AI in the future, they will surely lag behind and eventually be eliminated."

Lantao Lan recognizes that "most enterprises cannot possess the underlying research and development capabilities of large AI models like Baidu's." Therefore, Yingteng Education chooses to collaborate with Baidu's large AI model, both on the production side and in exploring applications on mobile apps.

Zhang Haoran, Vice President of Motiff Miaoduo, an enterprise-level AI-driven interface design tool, believes that "combining AI with professional tools will bring new models and opportunities to existing professional tools and production relationships."

Among them, a large amount of high-quality data is the key to improving the capabilities of large models. "The vast professional UI interfaces accumulated by Motiff Miaoduo represent a barrier to large model capabilities, enabling enhancements and activations in specific domains," said Zhang Haoran.

Wang Yunfeng, CTO of Beijing Zhidemai Technology Co., Ltd., found that AI is a double-edged sword. While it significantly enhances content production efficiency, it also exacerbates the proliferation of content.

Based on consumer behavior feedback data, more users are being more cautious during the purchasing process, fearing being misled by the deluge of low-quality information. On the other hand, users continue to spend money in their areas of interest.

Based on this, Zhidemai Technology helps users make informed purchasing decisions by personalized recommendations for products of interest and introducing its self-developed shopping assistant "Xiaozhi" to assist users in their research. "Baidu Baihe's heterogeneous computing architecture allows us to provide services to users at a reasonable cost," said Wang Yunfeng.

Liya Li, Founder and CEO of Renben Zhihui-Fenghuang FM, offered two keywords: "Collaborating with Hardware + AI" and "Dual-line Information Flow + Agent Distribution."

Liya Li observed that hardware access points are crucial. Many new AI hardware devices, such as earbuds, necklaces, and smart home appliances, are competing with dominant devices like smartphones and cars for access points. Fenghuang FM, the core application under Renben Zhihui, has jointly created a unique dual-line information flow model with Baidu Intelligent Cloud's Wenxin large model and various intelligent hardware devices. Based on the specific semantics and scenarios of thousands of news items daily, it provides a high-frequency news and knowledge linkage experience and actively explores the distribution and dissemination of strongly related knowledge, stories, content payments, product advertisements, and agents based on generative AI.

Zhongchao Shi, Chief Researcher at Lenovo Group and R&D Director of Lenovo Research's AI Lab, said that Lenovo has a large number of client-side devices and believes that the client side will be the primary access point in the future. "Last year, Lenovo and Baidu collaborated on both the client and cloud sides. Currently, there is a cloud-based model application on AIPC products, which is also in collaboration with Baidu."","I hope that the ideal of AI technology can be realized as soon as possible," Shi Zhongchao candidly admitted.

Chu Chang, Director of AI Platform Product Solutions at Baidu Intelligent Cloud, believes that "currently, large model technology is still in a stage where great efforts can lead to remarkable achievements, and the ceiling continues to rise."

Applications across various industries and industry partners have accumulated profound domain-specific knowledge and experience, with a deeper understanding of scenarios. Chu Chang said, "We will do our job well and empower our partners based on large models and toolchains." Baidu Intelligent Cloud's Qianfan large model platform can help enterprises use large model technology more "smoothly" and seize the initiative in the era of large models. It is reported that currently, on Baidu Intelligent Cloud's Qianfan large model platform, the daily call volume of Baidu Wenxin large model has exceeded 700 million, with over 30,000 fine-tuned large models and over 700,000 developed applications.

Looking back at the 30-year history of the Internet, it has been a rollercoaster ride. Initially navigating through the fog of speculation, it then experienced a golden decade of mass entrepreneurship and innovation, seamlessly integrating into everyone's work and life over the years.

Now that the Internet is embracing a wave of "intelligent transformation," new business stories are staged in turn, with exciting new applications emerging daily. In the new era of global AI competition, the pioneering spirit accumulated over 30 years of the Internet in China will continue to serve as the essence guiding industrial innovation.

*The lead image in this article is from the Interface News Image Library.