Although the path of commercialization of large models is relatively clear, and domestic manufacturers are also actively exploring, the commercialization of large models cannot be limited to the exploration of business models, but also lies in solving the underlying problems of large model development.
Author |Fighting.
Edit|Pi Ye.
Produced by |Industrialist.
Nowadays, the commercialization of large models is once again on the table.
It is a fact that the current large-scale model training requires strong computing power, especially for models with a large number of parameters, which is extremely expensive. For example, OpenAI's language model GPT-3 costs nearly 5 million US dollars, which is about 40 million yuan. Huge model training requires huge amounts of money to support.
After investing a huge amount of money, on the one hand, the company hopes to commercialize as soon as possible to solve the problem of follow-up R&D funds, and on the other hand, it also hopes to achieve the purpose of making money through commercialization.
Then the contradictions also followed, rapid commercialization, it is inevitable that some safety and ethical issues have been put aside for the time being, and the more real situation is that the development path after the rapid commercialization of large models is not much to think about, so we have seen that many large models are basically just a tasteFinally, it leads to the contradiction between commercialization and non-profit.
OpenAI's "Gong Dou" some time ago is a good example.
On November 18, OpenAI's management changed dramatically, and CEO Ultraman was fired. At this point, OpenAI's "Gong Dou" was staged.
In public reports, among OpenAI's six-member board, the fired Altman and Greg Brockman are inclined to accelerate commercialization to obtain more funding to support the computing power needs of AI modelsIndependent directors Tasha McCauley and Helen Toner are more concerned about AI security.
In short,One side is technology-led, pursuing model excellence with the goal of achieving general artificial intelligence;On the other hand, commercialization is the only way for the company to developIt is necessary to actively expand market applications to achieve profitability. As a result, one speculation is that Altman, who advocates commercialization, ran into Ilya Sutskever, who emphasizes AI technology and security attributes, and directly triggered a confrontation.
After repeated tug-of-war. On November 30, OpenAI announced the formation of a new initial board of directors, with Sam Altman returning to the role of CEO Mira Murati as chief technology officer. The winner of this "palace fight" seems to belong to the commercial side.
However, under the farce of "commercialization and non-profit dispute" caused by the world's top large-scale model companies, some questions have made people think deeply, that is, what dilemmas are the commercialization of large-scale models facing?How should large models be commercialized?
In the Chinese market, in addition to the commercial value of computing power that has been demonstrated, what other aspects can large model manufacturers try to commercialize?And, where has this road gone?
First, the commercialization of large models
In terms of the commercialization of large models, Internet manufacturers represented by Alibaba, Tencent, etc., currently have a clear commercialization prospect. This is not unrelated to its own large business system.
That is, Internet giants can integrate large models into existing products and services, such as library document assistant, **Qoi, Bing search engine, etcIncrease user stickiness and drive revenue growth. The main way is to embed generative AI as an auxiliary function and embed it into the original business as a value-added service.
The second is the subscription service, which is a monthly or pay-as-you-go subscription model that provides customers with ongoing access to large models. For example, OpenAI's ChatGPT, Waiting, Wenxin Yiyan, Alibaba's Tongyi Qianwen, etc. At present, domestic Wenxin Yiyan and others are also bringing some revenue to large-scale model applications through the subscription-based business model, but the charging intentions of other manufacturers are unclear.
In addition, the commercialization prospects represented by national team manufacturers such as Zhipu AI are also relatively clear. It is generally believed in the industry thatFor example, large domestic enterprises and central state-owned enterprises want to combine with large models, and Zhipu AI is an unavoidable option.
However, despite this, the commercialization of large models in China is still in its infancy, and the commercialization process is facing many challenges.
First of all, the development and application of large models require a lot of capital and time investment, and the return is often difficult. This has led many companies to hesitate in the commercialization process and miss out on market opportunities.
Secondly, the ethical and safety issues of large models also bring certain pressure to commercialization. For example, issues such as algorithmic bias and discrimination, data breaches, and abuse occur from time to time, which makes some companies wary of using large models. In addition, the commercialization of domestic large-scale models also faces problems such as market acceptance and application scenarios.
At present, the application needs of most enterprises are mainly concentrated in the fields of intelligent customer service, intelligent recommendation, and intelligent marketingApplications in other areas are still in the exploratory stage. This makes the commercialization process of large models relatively slow, and it is difficult to achieve large-scale development.
What is more noteworthy is that although China has made significant progress in the field of artificial intelligence, there is still a certain gap in domestic large-scale model technology compared with the international leading level. This puts domestic enterprises at a disadvantage in the international market competition, and it is difficult to extend to the sea and cross-border.
In addition, the commercialization of domestic large models also faces the problem of immature business models, such as how to charge, in terms of the form of computing power fees commonly adopted in China, this model seems to be consistent with the charging model of cloud computingAnd from the perspective of profit margins, this is obviously not a high-quality charging model.
For domestic large-scale model manufacturers, how to go on the road of commercialization has become an urgent problem to be solved.
2. MaaS, open source, and agent
The commercialization of large models should solve the problem of letting enterprises and users understand less about the principles, use the results more simply and directly, and let users return to value and solve their own business problems. In other words, it is the "all-in-one black box model" of the large model.
As a result, some of today's business models have become a gathering place for large-scale model players and entrepreneurs.
Among them,The MaaS pattern is the most common. In this mode, cloud vendors or scientific research institutions generally encapsulate large models, encapsulate inference capabilities on various tasks into a unified application interface, and provide services to the outside world.
Downstream enterprises can obtain these interfaces and embed them into existing applications and services according to their own business needs, so that the APIs of the large model can empower the entire program.
In this way, enterprises do not need to know too much about the technical details of the model, but can directly call the service as if it were a cloud capability. At present, large model manufacturers such as Wenxin, Tongyi, and Pangu are basically providing such services, such as Ali's magic community, paddle and so on.
In addition,Open-source modelIt is also an important way to commercialize large models, in which the source of computer programs and software is disclosed and distributed according to open source licenses.
Open source is a common software development model in the computer field, a large number of developers modify open source ** under the license of the agreement, and integrate it into the existing system to add new functions and features to the software and system.
In the open source mode, good results can be quickly shared, so that good results can quickly cultivate the community, and downstream users can use open source results to quickly build their own application systems. In China, Zhipu AI and Ali Tongyi are emphasizing the value of open source.
Open source itself is free, but when it comes to follow-up data training, data supervision, data fine-tuning, etc., it corresponds to a clearer charging model, which is equivalent to opening up well water, but being a shovel seller.
And then there isPlatform-as-a-Service model, i.e. no longer providing a single model API, but:Think of the big model as a technology in the platform serviceIntegrate into the AI platform and provide services to the outside world through a unified platform. In this model, enterprises build a platform that includes development tools, AI services, and processes, and the large model is only one component of the platform.
In the process of purchasing or using the platform, users can use the tools provided by the platform to develop and apply large models, which are integrated into their own systems, and users cannot obtain the capabilities of the models alone. By using the platform and tools, users gain the ability to develop with large models, and they pay for it.
For example, Wenxin has developed a large model of NLP CV cross-modal biological computing, and on this basis, many industry large models and large model suites have been launched. There are easy-DL, BML large model, large model API, Wenxin One Grid (AIGC) and so on.
There is another kindSoftware as a Servicemode. At present, large domestic factories, leading government enterprises and scientific research institutions are providing strong new infrastructureSmall and medium-sized vendors can develop their own SaaS services based on these infrastructures and provide them to enterprises and individuals. AI agent is the hottest large-scale model entrepreneurship path at present.
In addition, whether it is for AI leaders such as OpenAI and Meta, or for many small start-ups or tech geeksAI agent is also a topic that has to be talked about in commercialization today. Whether it is DingTalk, Feishu, or even others, they are all launching their own agent products.
If many of the aforementioned realizations are on the B side, there is a certain ceiling in its market and demand. Then, AI agent corresponds to the huge imagination on the C-side outside the B-end market. Not only the market itself, but also in the business value.
Today, a general consensus in the industry is:AI agent is the only way to realize the ultimate form of AGI (Artificial General Intelligence) in the futureMoreover, more and more people realize that large models can only show their true value when they enter thousands of households at the real application level, and AI agent is the best application form.
3. What is the difficulty in commercialization?
Overall, although the path to commercialization of large models has not been the best, the direction is clear. But clarity doesn't mean it can be landed. For players in the domestic large-scale model track, they still face many internal and external challenges.
In the early morning of November 7, OpenAI released several updates at the first developer conference, the new model GPT-4 Turbo, GPT Builder, and Assistant API.
Among them, the features of GPT Builder include that everyone and every business can customize their own GPT;Each unique GPT can be customized with its own commands, knowledge base, tools and actions, avatars, etc.;No development required, directly using natural language customization, you can even ask dalle3 to generate avatars for you;GPTS can be shared and used and enjoy a share similar to App Store.
This means that every business can create their own GPT agent.
The other is to update the Assistant API, which allows GPT to write ** for you and execute it automaticallyImplement the ability to call functions and tools through APIs, and extend the capabilities of AI.
This means that users can more easily build their own chatbots or AI assistants in their own ** or mobile applications through the Assistant API, greatly reducing the heavy workload of AI development.
A fact is,It is no longer satisfied with providing a basic large model, but hopes to become an AI OS platform in the AI era. This update has largely had a big impact on the sales model of AI Agent.
In the open source model, there are also bottlenecks in development. Taking Zhipu AI as an example, the current model parameters of Zhipu AI are mainly 6B, and the parameters are small. The reason for this is inseparable from the problem of insufficient funds. It is important to know that the larger the model parameters, the greater the computing power requirements. Although Zhipu AI has purchased a large number of A100 before, judging from its recent frequent and high financing, it still needs a lot of funds to support its continuous commercialization and R&D innovation.
There are also many problems in the implementation of the MaaS model. First of all, if the model effect is not satisfactory, the API will not be able to fully meet the user's conventional inference needs, so the model needs to be adjusted and optimized according to the specific situation, but tuning itself is a development with a threshold, and most enterprises do not have such capabilities or large model talentsThis makes it difficult to continue to contribute to the MaaS community.
Second, due to the relatively slow running speed of large models, it will be difficult to guarantee the response time and data quality of APIs when the number of inference requests or the amount of requested data increases significantly. For example, AIGC applications such as ChatGPT and Dalle2 tend to have a long response time, making it difficult to achieve large-scale applications and provide timely response experience in a short period of time.
Overall,The commercialization of the global large-scale model industry is still in the early stage of exploration.
On the one hand, although R&D institutions are quite mature in terms of large-scale model technology, they are not familiar enough with the landing scenario and have not yet formed a complete commercialization model. Therefore, they need to cooperate with downstream scenario enterprises to jointly build a large-scale business model.
On the other hand, most downstream scenario enterprises have not yet formed the basic concept and cognition of large models, and at the same time, they also lack the computing power required to support model fine-tuning, as well as the human resources and technical strength required for customization and secondary development of models.
In general, although the path of commercialization of large models is relatively clear, and domestic manufacturers are also actively exploring, the commercialization of large models can not only be limited to the exploration of business models, but also to solve the underlying problems of large model development.
Write at the end:
It is a fact that the real value of a large model lies in its ability to solve practical problems and create business value, and scenarios are the basis of the business model. For players of large model tracksHow to combine the large model with specific scenarios and implement them together 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 wide application potential.
However, without the right scenarios and applications, this tool can only stay at the theoretical level or in a laboratory setting. Only when it is successfully applied to various scenarios can it exert real business value.
copy.AI is a startup that uses GPT-3's large-scale language models to help businesses and individuals quickly generate high-quality content. Through an in-depth understanding of customer needs and market conditions, copyAI closely integrates the technical capabilities of GPT-3 with application scenarios such as marketing, advertising, and press releases, and realizes the transformation from technology to product. This "scene is king" strategy makes copyAI was able to stand out in a highly competitive market and become a high-profile startup.
Domestically, such an attempt may become the subject of the next stage.
List of high-quality authors