A few days ago, Fu Sheng and Zhu HE Tuber Xiaohu had a debate around Entrepreneurship". In addition to this incident, the issue of "how to commercialize large models" has also aroused keen attention from people in the industry. In this article, the author attempts to interpret and answer it, let’s take a look.

On the evening of June 26, Cheetah Mobile CEO Fu Sheng forwarded a WeChat article titled "Zhu Xiaohu is very unfriendly to startups" in Moments, with the comment "Half of the startups in Silicon Valley started aroundWe Investors can still be so ignorant and fearless.”

This repost quickly aroused a response from the person involved in the article, Zhu Xiaohu, managing director of Jinshajiang Venture Capital, saying that "the vast majority of startups are worthless in front of and gave examples of those who were most affected by in the segmented track. Two startups - Grammarly, which does automatic grammar correction, and Jasper, which usesto realize automatic marketing copy generation.

The debate finally ended with the two reaching a consensus on the point that "it is difficult for startups to get the opportunity to reinvent BAT (with the help of large models)." But outside of this debate, the issue of "how to commercialize large models" is the area of ​​greatest concern to the industry in addition to product progress.

After being silent for more than a month, OpenAI is also preparing for its next impact on the commercialization of large models: According to The Information, OpenAI CEO Sam Altman said in an internal meeting in June that it is planning to create a "big model". Model Application Store" is a new attempt in the commercial field of large models.

According to people familiar with the matter, the “application store” that OpenAI plans to launch will be a platform that allows OpenAI customers to put on their own customized large models and then conduct customized sales based on the actual needs of other companies.

 this is not the first time OpenAI has tried to make more "commercial": a month ago,

 OpenAI had launched a similar "App Store" project - At that time, OpenAI opened the third-party plug-in store, allowing users to download third-party plug-ins directly from it and integrate them into their own o expand to more usage scenarios.

However, this model did not go as OpenAI expected, allowing to set off a new wave of applications. So far, most plug-ins in the app store have been downloaded only tens of thousands of times. Several plug-ins developed by larger service providers, There are only a few hundred thousand downloads, and this model has not been truly promoted to most users.

This directly exposes OpenAI's real shortcomings in the field of large models: the commercial use of large models not only requires technology that is ahead of the times, but also requires a large model business operation model that is ahead of the times.

The main purpose of designing a large universal language model is to directly face billions of users around the world and answer tens of billions of various questions. It is not to solve a special industry problem in a certain professional field. The sources of required corpus are generally a wide range of public documents and Internet information. Internet information may not only contain errors, rumors, and biases, but also lack of accumulation of professional knowledge and industry data, resulting in insufficient industry pertinence and accuracy of the model. However, in industrial scenarios, users have high requirements for professional services provided by enterprises and low fault tolerance. Once a company provides incorrect information, it may cause huge legal liability or public relations crisis. There have been many similar incidents in the past.

This is the most important misalignment between and the needs of mainstream industry users, and it is also the reason why has not achieved more results in the commercialization of large models. ‍

 OpenAI encounters a bottleneck in its development of commercial large-scale models, it has to turn to Microsoft, its former partner and shareholder, to become a direct business partner. Competitors: If OpenAI will eventually open the "Large Model Application Store", it means that OpenAI will formally and deeply participate in the commercialization of large models.

It is also worth noting the dynamics of domestic large models: Although there is still a gap between the domestic large model industry and the first echelon such as OpenAI in the field of cutting-edge research, the research of the domestic large model industry in the field of commercialization is almost at the forefront of the world: According to According to the opinion of AI entrepreneur Ji Yichao (Peakji), in the future, the main application methods of large models will become ToC local operation and ToB privatized deployment. The former is a smaller "small model" deployed directly on a mobile phone/computer. Directly rely on the computing power of the device itself to complete some simple generative AI work.

This is also the reason why Google released four different volume versions of the PaLM 2 model at this year's I/O conference. Among them, the smallest large model, codenamed "Gecko", can be generated directly on the device's local side and run offline. This type of artificial intelligence does not occupy any cloud computing power. Relatively speaking, due to their small size, general general-purpose large models can carry tens of billions of parameters. These locally run large models may only need hundreds of millions of models to independently complete mobile phone applications including natural language understanding, speech recognition, image recognition, etc. Commonly used AI scenarios in .