AsianFin -- Most enterprises have begun integrating generative AI into their business operations in the past year, supported by Chinese and international large model providers, following the success of ChatGPT in establishing a technical pathway for general-purpose large models.

A recent IDC report shows that 70% of enterprises are optimistic about the potential of generative AI. By 2027, AI spending in the Asia-Pacific region is projected to reach $90.7 billion, with a compound annual growth rate of 28.9% from 2022 to 2027.

Compared to early 2023, the attitude towards large models has shifted from a "wait and see" attitude to an "all in" approach. The business model for general-purpose large models has now entered a new phase with tech companies engaging in price wars to battle for market share. The applications of large models have expanded from simple AI-generated content to long-text output, and some code assistants can now handle about 20% of a developer's work.

Public information reveals significant developments. Baidu's code assistant has written a quarter of its internal code; Baidu Search disclosed that 11% of its search results are AI-generated; Alibaba Cloud is fully implementing AI code writing, aiming for 20% of code to be generated by intelligent assistants in the future; NetEase Yanxuan disclosed that 50% of its product descriptions and marketing posters are AI-generated, with AI even participating in product design.

Alibaba Group on Tuesday slashed prices for several AI services by up to 97%, a move that further spark a price war in China's AI sector.

Following Alibaba's announcement of discounts on nine products based on its Tongyi Qianwen model, Baidu Cloud responded by offering free services utilizing its Ernie AI models. This came after ByteDance released last week that it would price its AI services—benchmarked against Ernie and Alibaba's Qwen—99% lower than typical Chinese industry rates.

Last Wednesday, TikTok's owner ByteDance launched a batch of large language models (LLMs) – the technology behind ChatGPT and other generative artificial intelligence (GenAI) services – that costs less than those from industry rivals.

These competitive moves mark the start of a price-driven battle within the AI sector, which is seeing substantial investment from startups and tech giants like Tencent. The influx of investment has led to the development of numerous AI models and an array of consumer and enterprise products, all vying for a large user base to accelerate AI advancements.

LLMs, which are trained on a vast amount of data, are revolutionizing GenAI applications such as chatbots, virtual assistants and advanced content-generating tools like Sora. GenAI are algorithms used to create new content, including audio, code, images, text, simulations and videos.

As of January, the number of government-approved LLMs and related AI applications on the mainland total more than 40. But at present, there are more than 200 China-developed LLMs in the market, which could lead other mainland providers to compete with Baidu in a price war.

Luo Xiaohui, CTO of Yeahka, a Chinese leading payment-based technology platform, led a 10-person AI Lab team to develop internal generative AI projects after ChatGPT emerged at the end of 2022. The company's main business in mobile payments and in-store e-commerce has guided their focus on applications like intelligent customer service, code assistants, and AI content creation tools. These applications have been progressively implemented since 2023.

Reflecting on the two-year journey of large model implementation, Luo's team exemplifies many enterprises eager to follow the trend. In an interview with AsianFin, he shared three key insights: First, the primary goal of large model implementation is to achieve "internal empowerment": The main consideration for implementing large models at Yeahka is "internal empowerment," focusing on enhancing research efficiency, marketing material creation, and customer service automation. Second, open-source vs. closed-source model is not a binary choice: Despite heated debates, open-source and closed-source models each have their strengths and suitable scenarios. Yeahka uses both, deploying open-source models on their private IDC and integrating closed-source models via API. Third, computing power bottleneck is manageable: Unlike many internet companies relying heavily on public clouds, Yeahka uses its self-built IDC private cloud. They have stockpiled Nvidia A100s and supplemented with A10s due to restrictions, ensuring adequate computing power for training and inference with regular GPUs.

Interview transcript edited by AsianFin for clarity and brevity:

AsianFin: What prompted your interest in generative AI research?

Luo: We started exploring generative AI at the beginning of 2023, deeply impressed by ChatGPT. As a company focused on applying large models, we aimed to uncover their value, starting with content generation needs in our in-store e-commerce business, such as creating marketing content, WeChat articles, and short videos.

AsianFin: Many companies are going "all in" on generative AI. How do you assess its strategic importance?

Luo: Any technology is useless if it can't find a practical application and generate value. We first identify the intersection of technology and our business needs. Next, we conduct a series of validations, effect testing, and optimization.

AsianFin: Has the introduction of generative AI affected organizational structure?

Luo: The AI Lab team is still in charge, but in the second phase of validation, the business team participates in evaluating the effectiveness, integrating it into business processes. For example, developing the image and text generation tool took about six months, with two months of technical development and three to four months of business team optimization.

AsianFin: Any challenges or lessons?

Luo: General-purpose models can have limitations in specific scenarios due to a lack of domain-specific knowledge. The best approach is to combine examples and local data. We integrated previous local knowledge graphs for better results.

AsianFin: How do you manage project timelines and challenges?

Luo: Different teams have different rhythms. We are given flexibility, and as we approach deployment, we align closely with business objectives. Balancing goals and resource allocation, such as choosing between ChatGPT-4.0 and 3.5 considering cost and response speed, is crucial.

AsianFin: What's your take on the open-source vs. closed-source debate?

Luo: Both have their advantages. Open-source is more suitable for deep integration with business scenarios, especially for companies needing private deployment. Closed-source models might fit companies with limited technical resources.

AsianFin: Have you used both open-source and closed-source models?

Luo: Yes, we use both routes.

AsianFin: Is computing power a significant challenge?

Luo: We mainly do localized learning and fine-tuning on top of foundational models, which our current computing power can handle adequately.

AsianFin: Are you using domestic GPUs?

Luo: No, we stockpiled Nvidia A100s before restrictions and supplemented with A10s afterward.

AsianFin: For AI applications like intelligent customer service and code assistants, do you develop in-house or use existing products?

Luo: We develop in-house. Standard products don't offer the customization and optimization we need, particularly for local data integration. Using external code assistants poses security risks we cannot accept.

AsianFin: How effective is your intelligent customer service after optimization?

Luo: Our automated customer service resolution rate improved from 50% to over 80%, significantly reducing costs and allowing the customer service team to handle more complex issues.

AsianFin: What about the code assistant's effectiveness?

Luo: Currently in trial, it fits local coding habits and is improving efficiency, with some developers having 30% of their code generated by the assistant.

AsianFin: Will you engage with industry-specific large models, like those in finance?

Luo: We are watching but haven't planned concrete steps. The specific needs of financial sectors like payments, banking, and insurance vary, and further exploration is needed for effective implementation.

AsianFin: How significant is your data asset accumulation?

Luo: As a cloud-native company using a private cloud, private deployment of large models is essential for us.

AsianFin: Why not primarily use public cloud?

Luo: While cloud adoption was beneficial, the need for autonomy, especially in the financial sector, makes private cloud a better choice.

AsianFin: What are your future plans for large models?

Luo: We plan to explore applications in metaverse gaming assets, cross-border issues in our global strategy, and internal efficiency improvements. We might also develop market-ready products in the future.