July 22 session
Centralized large model: refers to the large model represented by chatgpt, which is closed source and relies on API revenue.
Decentralized large model: refers to large models that are open source and independently deployable, represented by llama and Baichuan.
Zhang Peng
①The large model is not a centralized operating system, but more like a computer. In the future, each individual may have more than one large model. In the long run, the centralized large model only accounts for 20% of the value.
②The large model of industry centralization is likely to be a false proposition. Enterprises will not be willing to share their knowhow, but are more inclined to enhance the company's competitiveness through closed-source large models.
③There are three kinds of innovations: first-hand innovation, offensive innovation, and revolutionary innovation. For companies with strong C-end stickiness, they are not in a hurry to release large-scale models. At that time, it is enough to add large-scale model functions to their own products. For example, Byte and Tencent.
④We are at a critical moment similar to the launch of the iPhone in 2007, when only a few people could see the huge potential of the iPhone and Apple. Mobile Internet companies began to explode in 2011, with an interval of four years. The time interval between a new technology and the outbreak of a new era is only getting shorter and shorter.
⑤The most valuable part of innovation often comes from non-consensus.
Wang Xiaochuan
① In the future world, 80% of the large models are open source, and only 20% are closed source, similar to Apple and Android.
② The evaluation score of the Baichuan model in the Chinese field far exceeds that of the English model.
③ Of the 9 technological innovation points of llama2, 6 are already being worked on by Baichuan, and 2 bring new inspiration.
④ The Baichuan model is completely self-developed. Due to its good effect, investors once mistakenly thought it was a llama casing.
⑤The large model is very dependent on high-quality data, and low-quality data will make the large model stupid.
⑥Search technology has great advantages in extracting high-quality data, model training, and experiment iteration.
⑦The cost of investing in large-scale model training is extremely high. If there is no substantive application, this investment may lead to a collapse of investment. Up to GPT4, there has not been a super application based on a large model.
⑧OpenAI's data flywheel has been falsified. The real core lies in the preparation and matching of high-quality data. The data of llama is precise but less. The matter of data matching has been scientifically implemented within openai.
July 23 session
Zhou Hongyi
① In the future, there will not be only a few large models monopolizing, and large models will be everywhere.
② Open source brings technological equal rights.
③The industrial revolution can be triggered only when large models enter thousands of households. Analogous to the process from the earliest computers to personal computers.
④ 80% of the opportunities are in the enterprise market.
Many companies are fascinated by large models, but they have concerns. I hope that the large model can be combined according to my own business, but I don't want the data to be leaked.
⑤ Seven major pain points of using large models in the enterprise market:
Pain point 1: Ask him everything, but his cognition is relatively shallow.
Pain point 2: I don’t understand the enterprise, and I don’t have the internal knowledge of the enterprise.
There is knowhow inside the enterprise, so it is definitely not willing to train it into the public large model.
The United States is dominated by public clouds. China is very concerned about where servers are placed.
Pain point 3: hidden dangers of data.
Pain point 4: vague knowledge and hallucinations. The content is not credible
Pain point 5: The problem of not updating knowledge in time.
Pain point 6: High cost.
Pain point 7: I am very concerned about the ownership of large model data. Security is paramount.
For example, the data of Sinopec cannot be given to PetroChina.
⑥ The vertical model is the Cotai Strip, piercing the market with a needle.
⑦ The most commonly used functions of large models are knowledge quizzes and content creation.
First find the most detailed vertical scene to apply these two functions.
⑧Four best demand types: upper, lower, inner, outer
Top: information decision-making, public opinion analysis
Next: Enterprise knowledge search, new employee training
Inside: Office Collaboration, Office Generation
Outside: intelligent customer service
These scenarios are more acceptable to enterprise customers and can be implemented relatively quickly.
⑨HIM、WHO（ChatUI）、GUI
LUI is still for enthusiasts, and the threshold is still relatively high. Most users don't research prompt words to this extent.
CUI defines some functions as digital people, digital employees, digital consultants, and digital experts. When people face a chat interface, it is difficult to think of prompt words. But when facing a legal expert, he will never ask some questions that are not related to law.
GUI, the traditional interface will still exist. will still be in parallel.
monica (Xiao Hong) & instant design (Li Guorui)
①GPT is the brain, and the application is the data. The value of the application lies in obtaining data that the brain does not have. Products that are directly shelled are of little significance.
②Browser plug-in is a very good form to obtain GPT without data.
③Browser plug-ins can accelerate the popularization of AI technology faster. There are many websites that have not yet been upgraded with AI, but AI capabilities can be used through browser plug-ins.
④monica now has 1.5 million users.
⑤ Most users still use the search and question-and-answer functions.
⑥ Therefore, the next step is to make many different robots to undertake questions and answers in different fields.
⑦The real-time design enables Party A, who was unable to generate the front-end page before, to generate the first draft, which improves the communication efficiency of the designer company. Reduced the design cost of Party A's company.
Fu Sheng
①The knowledge on the Internet is only a small part of the total knowledge, and the real and practical knowledge is in each company. It is not available on the Internet.
② Open source is a gold mine. Such as llama2 and Baichuan Smart.
③The knowledge base is not so easy to do. We must grasp the details, and the devil hides in the details.
④ prompt is very important.
⑤The present is the future of the past, and the present is not the past of the future.