Technology Empowerment, Localization Accelerates Breaking Waves" roundtable forum.
On November 29th, the 36 Krypton Wise Conference "Gold Rush in the Waves - Capital Market Conference" was held in Beijing, with the theme of "Technology Empowerment, Localization Accelerates Breaking the Waves" roundtable forum, Zhu Yun, chief analyst of West China ** overseas, Su Peng, vice president of NetEase Youdao, and Huang Ying, marketing director of Black Sesame Intelligence, conducted a roundtable forum on the investment and current challenges of hard technology.
The following is a summary of the content of the dialogue, which has been deleted.
"Now the progress of artificial intelligence in China is not a cyclical problem at the technical level, but there is still a certain gap in the underlying technology in China. ”
——Zhu Yun, chief overseas analyst of West China
This year, whether it is the heat of the primary market or the performance of the secondary market, it is indeed relatively unexpected, and the performance of the entire secondary market this year is not very good.
But we should look at it separately from foreign and domestic. We are concerned that Microsoft in the United States is a record high, so they are not in a state of decline, but there may be some reasons in China, after the large model reaches a certain stage, there are not many applications that make everyone shine.
Let's look at these leading technology companies in the United States, for example, looking at the data, the recent chatgpt, which cannot be used in China, may not be felt by everyone, but from the latest chatgpt daily and weekly activity data, the recent login data has hit a record high, close to the new high in March and April. So it's just that everyone thinks that (the big model) doesn't seem to be very good in China, because of the poor information, and the whole performance overseas is still great.
I think that talking about artificial intelligence now is not a cyclical problem at the technical level, but it may be that there is a certain gap in the underlying technology of artificial intelligence itself. The C-end application, which is closely related to everyone, is not very clear.
In our previous report "Review and Enlightenment of AIGC Non-general Inflation** in the United States", we mentioned process management companies and large model empowerment companies, mainly some computer companies. We did a forum in April and May this year, and the feedback they gave me at that time was that there would be artificial intelligence software companies in the financial industry at the end of this year, and there would be small models of some companies in the pharmaceutical industry and the application of industry data. But now it feels like it's not as good as it could be.
It is recommended that you take a look at the progress of overseas artificial intelligence development, which will be better than what you see so far. The trend of artificial intelligence has not changed, although Mr. Wang also mentioned that there have been some farces recently, but I think the whole trend has not changed.
In any case, the progress of science and technology as a whole is positive. If you want to make an investment, you should still grasp the main line, but maybe the better development at the moment is overseas.
"The difficulty of the industry vertical model is that the data quality requirements are very high. ”
——Su Peng, vice president of NetEase Youdao
The relatively difficult part of the training of large vertical models in the industry is the requirements for data, and the quality requirements will be higher. We need the model to be more like the output of a professional. A simple analogy, like the company's induction training, is equivalent to turning him from a generalist to a specialist, and on the basis of the specialty, other fields do not need to be too sophisticated, and you need to be very professional in the vertical field.
This requires a high level of data quality as a whole. In the training process, whether it is our best teacher or the best tutor, we have to participate in the whole training process, and they become the very important output to provide the basis for data training and organize the data. He wants to output a set of knowledge points, educational principles, and educational methods to the model, and form a highly executable trainable text, which will be converted into training data by technicians, and then the data will enter the model to start training.
It seems to be a technical process, but the whole training process is still a process of data collection, labeling, sorting, output and then training, which requires the participation of two parts, in addition to technical personnel, there are also industry professionals to understand the needs of data personnel, what kind of training level is needed, what level of text needs to be detailed, and opinions should be expressed for technical personnel to understand.
Whether it is the application of a large model in a professional field or the professional training of a vertical model, this is important in the professional field and requires mutual understanding between the two parties. Technicians need to understand the characteristics of the industry and be able to precisely target training needs, or data needs.
For the person providing the data, he should have his own understanding of the technology industry, know how to respond to the professional needs of professionals, how to make his data more in line with the standardized data collection process, we understand, this is the most challenging part of the whole process. Whether at the content or technical level, Youdao has spent a lot of energy and resources on exploration, and scenario landing is our primary consideration.
"It's just a chip, but it's not enough at all. If the car company can use it well, it also needs the support of software. ”
——Huang Ying, Marketing Director of Black Sesame Intelligence
First of all, we have now fully entered the era of smart cars. In the era of smart cars, it can be seen in the internal strategies of car companies and Tier 1 companies that the development trend of the electronic and electrical architecture of the whole vehicle is from distributed to first-class.
Secondly, in fact, this year, everyone often mentions cost reduction, because everyone wants to make money, and cost reduction is also very important. Based on such a large background, if you want to support the chip of cross-domain computing, you must support how it is flexible and better to use in the design of the chip, so we launched the first chip C1200 in the Wudang series, which has been unveiled at the Guangzhou Auto Show, and can now send samples to customers.
Cabin and driver integration is also a hot word in the industry, and it generally requires two chips or more to achieve this function. If there are more domains, there may be more chips. The C1200 of the Wudang series can support two domains or even four domains with one chip. It not only supports multi-domain convergence, but also supports the application of many functions and scenarios. Because it can be done with only one chip, the power consumption will be greatly reduced, and the cost performance will be greatly enhanced. The C1200 is a completely innovative architectural design.
Of course, it's just a chip, but it's not enough at all. If the car company can use it well, it also needs the support of software. In addition to chips, Black Sesame Intelligence provides complete solutions for car companies. Middleware, toolchains, software, algorithms, etc. Together with many partners, we have developed more than 30 domain controllers, which can be applied to different scenarios and fields.
Stay tuned for more information.