How Should China's Large AI Models Be Commercialized? AI Applications, MaaS or Agent
-
Now, the commercialization of large models is once again on the table.
One fact is that current large model training requires immense computational power, especially for models with high parameter counts, which comes at a staggering cost. For example, OpenAI's language model GPT-3 cost nearly $5 million, approximately 40 million RMB. Training such massive models demands enormous financial support.
After investing huge sums, companies hope to commercialize quickly to address follow-up R&D funding issues while also aiming to generate profits through commercialization.
However, conflicts arise inevitably. Rapid commercialization often sidelines safety and ethical concerns. More critically, there is little deep thinking about the development path post-commercialization, leading many large models to merely scratch the surface. Ultimately, this triggers a conflict between commercialization and non-profit objectives.
The recent 'power struggle' at OpenAI serves as a prime example.
On November 18th, OpenAI's management underwent a major shakeup when CEO Sam Altman was dismissed, marking the beginning of the company's internal power struggle.
According to public reports, within OpenAI's six-member board, the ousted Altman and Greg Brockman leaned towards accelerating commercialization to secure more funding for supporting the computational demands of AI models. In contrast, independent directors Tasha McCauley and Helen Toner prioritized AI safety.
In short, one faction is technology-driven, pursuing model excellence with the goal of achieving artificial general intelligence; the other believes commercialization is an inevitable path for the company's growth, advocating for aggressive market expansion to achieve profitability. This has led to speculation that Altman, who championed commercialization, clashed directly with Ilya Sutskever, who emphasized AI technology and safety attributes, sparking the conflict.
After repeated back-and-forth, OpenAI announced a new initial board on November 30, with Sam Altman returning as CEO and Mira Murati as CTO. The winner of this 'power struggle' seems to be the commercial faction.
But beneath this farce of 'commercialization vs. non-profit' sparked by the world's top large model company, some questions arise: What challenges does the commercialization of large models face? How should large models be commercialized?
In the Chinese market, beyond the already demonstrated commercial value in computing power, what other areas can large model companies explore for commercialization? And how far has this path progressed?
In terms of large model commercialization, internet giants like Baidu, Alibaba, and Tencent currently have relatively clear commercialization prospects. This is closely tied to their extensive business ecosystems.
Internet giants can integrate large language models into existing products and services, such as Baidu Wenku Document Assistant, Taobao Ask, and Bing search engine, to increase user engagement and drive revenue growth. The primary method is to incorporate generative AI as auxiliary features into existing businesses, treating it as a value-added service.
Another approach is subscription services, which provide continuous access to large models through monthly or usage-based billing models. Examples include OpenAI's ChatGPT, Baidu's ERNIE Bot, and Alibaba's Tongyi Qianwen. Currently, domestic products like ERNIE Bot are generating some revenue through subscription-based business models, while other manufacturers' pricing intentions remain unclear.
Additionally, state-backed manufacturers like Zhipu AI demonstrate clear commercialization prospects. The industry generally agrees that for large domestic enterprises and state-owned companies seeking to integrate with large models, Zhipu AI is an unavoidable option.
However, despite this, the commercialization of domestic large models is still in its infancy, facing numerous challenges in the commercialization process.
First, the development and application of large models require significant financial and time investments, while the returns are often unpredictable. This leads many enterprises to hesitate in the commercialization process, missing market opportunities.
Second, ethical and safety issues related to large models also bring certain pressures to commercialization. For example, problems such as algorithmic bias and discrimination, data leaks, and misuse occur from time to time, which makes some companies cautious in applying large models. Additionally, domestic large model commercialization also faces issues such as market acceptance and application scenarios.
Currently, most enterprises' application needs are mainly concentrated in areas such as intelligent customer service, intelligent recommendations, and intelligent marketing, while applications in other fields are still in the exploratory stage. This makes the commercialization process of large models relatively slow and difficult to achieve large-scale development.
More notably, despite significant progress in the field of artificial intelligence domestically, there remains a certain gap between China's large model technology and international leading standards. This puts domestic companies at a disadvantage in international market competition, making it difficult to expand into overseas and cross-border markets.
Furthermore, the commercialization of domestic large models faces challenges due to immature business models. For instance, the prevalent compute-based pricing model in China appears identical to cloud computing pricing structures. From a profit margin perspective, this is clearly not an optimal pricing strategy.
For domestic large model developers, determining the path to commercialization has become an urgent issue to address.
The commercialization of large models should aim to help enterprises and users focus less on understanding the principles and more on directly utilizing the outcomes, enabling users to return to value and solving their own business problems. In other words, it's about creating an 'integrated black-box model' for large models.
As a result, certain business models have now become gathering points for players and entrepreneurs in the large model space.
Among these, the MaaS (Model as a Service) model is the most common one. Under this model, cloud providers or research institutions encapsulate large models, packaging their inference capabilities across various tasks into unified application programming interfaces (APIs) to provide external services. Although APIs are offered, the essence is the invocation of models.
Downstream enterprises can access these interfaces and, according to their business needs, call the services to embed them into existing applications and services, empowering the entire program with the capabilities of large model APIs.
This approach allows enterprises to utilize services without needing in-depth understanding of the model's technical details, similar to calling cloud capabilities. Currently, major model providers like Wenxin, Tongyi, and Pangu are offering such services, exemplified by Alibaba's ModelScope community and Baidu's PaddlePaddle.
Additionally, the open-source model serves as another significant approach to large model commercialization. Under this model, computer programs and software source code are made publicly available and distributed according to open-source licenses.
Open source has become a prevalent software development paradigm in the computing field. Numerous developers modify open-source code under license agreements and integrate it into existing systems to enhance software functionality and features.
Under the open-source model, valuable achievements can be rapidly shared, fostering community growth. Downstream users can leverage these open-source resources to quickly build their own application systems. In China, companies like Zhipu AI and Alibaba's Tongyi are emphasizing the value of open source.
While open source itself is free, subsequent processes like data training, supervision, and fine-tuning correspond to clearer monetization models—akin to offering free well water while selling shovels.
Another approach is the Platform-as-a-Service model, which shifts from providing standalone model APIs to treating large models as one technical component within platform services, integrated into AI platforms for unified service delivery. In this model, enterprises build comprehensive platforms encompassing development tools, AI services, and streamlined workflows, where large models serve as just one element of the broader platform.
During the process of purchasing or using the platform, users can leverage the tools provided by the platform to develop and apply large models, integrating them into their own systems. However, users cannot obtain the model's capabilities separately. By using the platform and tools, users gain the ability to develop with large models and thus pay for this service.
For example, the Wenxin large model has evolved into NLP/CV/cross-modal/biological computing large models, and on this basis, numerous industry-specific large models and large model suites have been launched. Further up, there are Easy-DL, BML large models, large model APIs, and Wenxin Yige (AIGC), among others.
There is also a Software as a Service model. Currently, major domestic companies, leading government enterprises, and research institutions are providing robust new infrastructure. Small and medium-sized vendors can develop their own SaaS services based on this infrastructure and offer them to businesses and individuals. AI Agent is currently a highly sought-after path for large model entrepreneurship.
Moreover, whether for leading AI companies like OpenAI and Meta or numerous small startups and tech enthusiasts, AI Agent is an unavoidable topic in today's commercialization. From DingTalk and Feishu to Baidu, everyone is launching their own Agent products.
If the previously mentioned monetization efforts were primarily focused on the B2B market, which has a certain ceiling in terms of demand and market size, then AI Agent represents immense potential beyond the B2B market—specifically in the B2C sector. This applies not only to the market itself but also to its commercial value.
Today, a widely accepted industry consensus is that AI Agent is the inevitable path toward achieving the ultimate form of AGI (Artificial General Intelligence). Moreover, more and more people recognize that large-scale models can only demonstrate their true value when they reach households at the application level, and AI Agent is the best form for such applications.
Overall, while the commercialization path for large models hasn't achieved perfection, the direction is clear. However, clarity doesn't guarantee successful implementation. Players in the domestic large model sector still face numerous internal and external challenges.
In the early hours of November 7th, OpenAI announced several updates at its inaugural developer conference, including the new GPT-4 Turbo model, GPT Builder, and Assistant API.
Among these, GPT Builder features include: every individual/enterprise can customize their own GPT; each unique GPT can be tailored with specific instructions, knowledge base, tools, actions, and avatars; no development required - simply use natural language for customization (you can even have DALL·E 3 generate your avatar); GPTs can be shared and used, with revenue sharing similar to the App Store model.
This means that every individual/enterprise can create their own GPT/Agent online.
Another update to the Assistant API allows GPT to help write code and execute it automatically through the API. This API enables function/tool calling capabilities, expanding AI's potential.
This means users can more easily build their own ChatBot or AI assistant on their websites or mobile applications using the Assistant API, significantly reducing the heavy workload of AI development.
A fact is that it is no longer satisfied with providing basic large models but aims to become the AI OS platform of the AI era. This update has largely impacted the sales model of AI Agents.
In terms of the open-source model, there are also development bottlenecks. Taking Zhipu AI as an example, the open-source model parameters are mainly 6B, which is relatively small. The reason lies in the challenge of insufficient funding. Larger model parameters mean greater computing power demands. Although Zhipu AI purchased a large number of A100s earlier, its recent frequent and substantial financing indicates that it still requires significant funding to sustain continuous commercialization and R&D innovation.
The MaaS model also presents several practical challenges. Firstly, if the model's performance is unsatisfactory, the API may fail to meet users' routine inference needs, necessitating adjustments and optimizations tailored to specific scenarios. However, fine-tuning is a development task with a high barrier to entry, and most companies lack the necessary capabilities or expertise in large models, making it difficult to sustain active contributions to the MaaS community.
Secondly, due to the relatively slow processing speed of large models, when the volume of inference requests or the size of request data increases significantly, the API's response time and data quality become hard to guarantee. For instance, AIGC applications like ChatGPT and DALLE2 often exhibit long actual response times, making it challenging to achieve large-scale adoption and provide timely user experiences in the short term.
Overall, the commercialization of the global large model industry remains in its early exploratory stages.
On one hand, although research institutions have achieved considerable maturity in large model technology, they are not yet familiar enough with practical application scenarios and have not established a comprehensive commercialization model. Therefore, they need to collaborate with downstream scenario enterprises to jointly build a business model for large models.
On the other hand, most downstream scenario enterprises have not yet formed a basic understanding and awareness of large models. Additionally, they lack the computing power required for model fine-tuning, as well as the human resources and technical capabilities needed for model customization and secondary development.
Overall, while the path to commercializing large models is relatively clear and domestic manufacturers are actively exploring it, the commercialization of large models should not be limited to experimenting with business models. More importantly, it lies in solving the underlying issues in the development of large models.
A fact is that the true value of large models lies in their ability to solve real-world problems and create business value, with scenarios being the foundation of business models. For players in the large model field, how to integrate large models with specific scenarios and achieve successful implementation is the essence of commercialization.
Taking OpenAI's GPT-3 as an example, this language model has attracted global attention with its strong generative capabilities and broad application potential.
However, without suitable scenarios and applications, this tool can only remain at the theoretical level or in laboratory environments. Only when it is successfully applied in various scenarios can it unleash its true commercial value.
Copy.ai is a startup that uses GPT-3's large-scale language model to help businesses and individuals quickly generate high-quality content. By deeply understanding customer needs and market conditions, Copy.ai tightly integrates GPT-3's technical capabilities with application scenarios such as marketing, advertising, and press releases, achieving the transition from technology to product. This 'scenario-first' strategy has allowed Copy.ai to stand out in a highly competitive market and become a highly regarded startup.
In China, such attempts may become the theme of the next phase.