In addition to Weibo, there is also WeChat
Please pay attention
WeChat public account
Shulou
2024-11-13 Update From: SLTechnology News&Howtos shulou NAV: SLTechnology News&Howtos > IT Information >
Share
Shulou(Shulou.com)12/24 Report--
According to the news on the afternoon of December 21, the annual ceremony of the "2023 Science and Technology list" jointly sponsored by Sina Financial client and Sina Science and Technology opened today. The theme of this year's activity is "Zhiyong Leap". Wu Hequan, academician of the Chinese Academy of Engineering, delivered a speech entitled "getting started with large models and improving Mathematical Intelligence".
Wu Hequan said that with the development of artificial intelligence, the Internet has rapidly entered an intelligent era. A year ago, ChatGPT gave everyone an eye. Of course, ChatGPT is still a basic model, and it is still the second generation of artificial intelligence. we hope that it will be universal in the future, at least in three aspects: first, a model can not only be used for one task, but can adapt to multi-task. Second, it can not be limited to one kind of modal input, multi-modal input is desired. Third, reasoning can be very accurate in the future.
He believes that the emergence of large models will bring good technical support for the digital transformation and high-quality development of our industry. Now the problem is how we can integrate more closely with our industry when we make large models. "I have noticed that several well-known units in China that build large models, although they have some conversations and chats that provide corpus, in fact, pure dialogue and chat is definitely not a rigid demand, and it is very difficult to form a business model."
In addition to the application scenarios in industry, Wu Hequan also gives an example. now the large model is on the mobile phone, and the mobile phone can already train more than 10 billion parameters. As a reasoning application, some people have already achieved 13 billion parameters. It is estimated that 15 billion may also be made into the mobile phone by next year. The landing of such a large model will lead to a new round of innovation and lower the threshold for users to generate 3D video. Our mobile phones were generally replaced in a year and a half many years ago, but there is no such demand in recent years. In the future, more practical promotion depends on user-generated content, and the large model falls on the mobile phone, which can improve the level of content generated for users. We are not only consumption, but also health, pension, guardianship, education, the most valuable is industrial applications, you can fall on mobile phones, you can also fall on robots, industrial modules, which can bring new leaps.
Citing IDC data, he predicts that more than 50% of the terminal processors in the Chinese market will have AI engines by 2026, which will bring China's Internet industry out of the trough and usher in an exciting blowout.
The following is a transcript of the speech: good afternoon, experts and leaders. The topic of my speech is getting started with large models and improving mathematical intelligence. In April next year, China will usher in the 30th anniversary of its successful entry into the Internet. We can recall that great changes have taken place in the past 30 years. China's Internet has entered its twenties.
It is found that the mobility of our entire Internet is developing rapidly, especially 4G has led to the rapid popularity of mobile Internet, and now 5G has driven the development of industrial Internet.
From the initial point-to-point connection for general consumer customers, the Internet has now become a platform, from point-to-point platform, cloud platform, social platform, live broadcast platform, industrial platform, especially in recent years, with the development of artificial intelligence, the Internet has rapidly entered an intelligent era.
Just now we talked about the big model. In fact, there was artificial intelligence research more than 60 years ago. In 1956, it was in the academic circles, and not many people knew about it. What is known to everyone is that the IBM Deep Blue computer defeated the chess master in 1997, and after a while, everyone felt nothing. Alpha defeated the go master in 2016 and found that humans could not play chess but computers.
Playing chess is not a rigid demand, nor does it have much impact on social life. At the end of last year, a year ago, ChatGPT gave everyone a bright eye. of course, ChatGPT is still a basic model and the second generation of artificial intelligence. we hope that it will be universal in the future, at least in three aspects:
First, a model can not only be for one task, but can adapt to multi-task.
Second, it can not be limited to one kind of modal input, multi-modal input is desired.
Third, reasoning can be very accurate in the future.
The large model also has many layers and many nodes, and we begin to correspond to a certain task, and we don't know which path is the most accurate, but we can know which node through large-scale training, repeated iterations, trial and error. What is the total probability that should be passed? these are the parameters. Obviously, the more parameters, the finer the decomposition. Now, from ChatGPT1 in 2018 to ChatGPT4 at the beginning of this year, the parameters have increased 10, 000 times, and of course, the corresponding training data and the number of calling GPU cards have also increased accordingly.
Now let's talk about big models. Just now the dialogue guests also mentioned that we have 188 big models in China. These big models are basically developed by Internet companies and IT enterprises. The threshold for this kind of basic large model is still relatively high. At present, there are very few enterprises in vertical industries. Large enterprises build large models themselves and make basic big models. As Dean Lin said just now, it has no way to be used in industry. The basic corpus is not industrial prediction. There is not so much data in industry, and it is not easy to find such a large amount of data for training. The basic large model can not fall to the ground on the node, what should we do? You have to cooperate with the industry, and there are two ways to cooperate with the industry. one is to send the enterprise's data to the provider of the basic large model, and then ask them to help add the industry data for fine-tuning. Will my data be leaked? Technically, it depends entirely on the basic large model.
There is also a way to give the model trained by the basic large model to the enterprise, and the enterprise itself adds its own data fine-tuning. Here, the technical level of the enterprise is relatively high. In addition, the basic training is taught by one teacher. When it comes to enterprise training, it is another teacher. Will there be any inconsistency between the two teachers? there may be no way to accept it in the end.
It is still difficult to cooperate with the industry, especially for most small and medium-sized enterprises, it is even more difficult to connect to the large model. We hope that we can turn the large model into a simple module between the cloud platform PAAS and SAAS, so that we can access this model module through a simple interface. We also need to configure some low-code development software accordingly, which can be dragged by the mouse to provide opportunities for enterprise basic scene access and fine-tuning. If we really do this, we will be able to use this model when the enterprise will go to the cloud in the future. I model small and medium-sized enterprises in this way, which I call the big model of the scene, and it is also aimed at specific applications.
The emergence of large models will bring good technical support for the digital transformation and high-quality development of our industry. Now the problem is still, how can we work as a basic large model side to integrate more closely with our industry? I have noticed that several well-known domestic units that build large models, although they have some dialogues and chats that provide corpus, in fact, pure dialogue and chat is definitely not a rigid demand, and it is difficult to form a business model. Domestic units that make some large models are aimed at the industry. For example, Baidu should cooperate with Geely to do intelligent customer service, and cooperate with the State Grid to do distributed power grid dispatching. Baidu also does effective analysis of MLA vaccine sequences. During the COVID-19 epidemic, inactivated vaccines were widely used in China, while MLA was used in the United States. There are many sequences, and not all of them are effective against COVID-19. It is still difficult to find the best sequence. It is said that it takes 10 billion years to exceed one second, but now a better vaccine can be selected by using a large model, which is not necessarily the best. Baidu has been published in a magazine and has been recognized. I think from these aspects of intervention, these aspects have not directly entered the manufacturing production line.
The Huawei Pangu model is mainly aimed at the manufacturing industry. it aims at the understanding of the needs in the manufacturing industry, the generation of documents, the programming of industrial software, the reading of drawings, and our supply chain management. We can also see that these are also on the periphery of the production line. it's really not at the core of the industry.
Tencent has a micro-build low-code platform, focusing on small and medium-sized websites, website development and so on.
Ali has a general meaning, a lot of training parameters, can support 8K to the above window, he can do chat dialogue, the length of your input also reflects the ability of the large model. Ali can enter about 8K.
The big model for the manufacturing field is Haier, which is a manufacturing industry. through Haier's own production of household appliances, he basically mastered the production process of household appliances. Haier big model has not been promoted in the household appliance industry, why? Others are his competitors, but the Haier model has been extended to the clothing industry, the automobile industry, and to these places.
With large models and the development of primary artificial intelligence, it also gives more opportunities for small and medium-sized enterprises in the society. At this time, there are a number of platform enterprises for more small and medium-sized enterprises. For example, there is an enterprise in Guangzhou that does modeling of clothing design and management of clothing factories. A large number of garment factories only have a large number of female sewing workers, no skills, the introduction of Guangzhou Zhijing software, so that the production management to a very good level.
Shandong Orange Cloud, which was originally a design tool software rental company, many enterprises use tool software, their own tool software is too expensive, use time is not much, rent, rent method can save money. Later, urban operation developed into a design undertaking and subcontracting platform, where many enterprises issued some requirements. He decomposed the design requirements, then invited tenders, and finally integrated the completed results through it. Now it is open to more than 50,000 small and medium-sized enterprises.
Shenzhen has a cloud technology, there are some enterprises need to order, need some products, do not know where to order, where to release. There are many enterprises that should bid on this, and can match 10 billion of the transactions in half a year.
There is a company in Guangzhou, mainly engaged in women's clothing export. It uses the clothing processing ability of the Pearl River Delta and international rapid logistics capabilities, from brand, design, fabric, procurement, sales, finance, insurance and so on. It is now the most important link in mobile shopping in 54 countries in the world. He is about to be listed and is valued at more than 100 billion US dollars. Jiangsu has a Zhiyun Tiangong, which is a virtual factory. Sany heavy Industry is the supply chain management platform. Sany works as the leader, connecting more than 200 upstream and downstream enterprises in the supply chain. The most important thing is to achieve zero inventory or less inventory. Greatly improve efficiency.
Now most of the big models are done in the Big Intelligence Center, which is super-calculated. Now a new one has come out, put the big model on the mobile phone, now the mobile phone can train more than 10 billion parameters, as a reasoning application, some people have achieved 13 billion parameters. It is estimated that 15 billion may also be made into the mobile phone by next year.
Some people say that there are only more than 13 billion parameters, what are the benefits of doing it on the mobile phone? in the future, the large model training can be offline, so the cost is low, there is no need for intelligent calculation, the super computing center, and the time delay is also low. Now there is a company in the United States, Aizip, in order to make a large model on a mobile phone, it needs to do some model compression work, and the mobile phone chip also needs to improve the file. To do the model compression work, we should quantify the compression and do it again. The company says small models can be copied from large models and can be landed on mobile phones.
Simultaneous interpretation, we call each other is a foreigner, he speaks English, I listen to Chinese, if it is a video, it can also help you match your mouth. We can have conversations with deaf-mutes, sign language and Braille interpreters. When you write a song, you hum a few paragraphs, and then I will continue it for you.
Search, in the past to be very accurate, now there is no need to be accurate, a vague word can also find out what you want to search. Of course, you can communicate with each other on mobile phone, tablet, PC and TV in the future.
In a word, a 32-year-old female conservationist explored the jungle and gave you this picture with a kind smile. The middle photo was only taken a little bit, but now it has been extended, maybe you only have a bust, and now it may become a full-length photo.
We now have front and back shots on our mobile phones, and now we can use them at the same time to embed your front photos into the back, and of course we have to adjust the light. This is a synthesis of selfies.
The landing of such a large model will lead to a new round of innovation and lower the threshold for users to generate 3D video. Our mobile phones were generally replaced in a year and a half many years ago, but there is no such demand in recent years. In the future, more practical promotion depends on user-generated content, and the large model falls on the mobile phone, which can improve the level of content generated for users. We are not only consumption, but also health, pension, guardianship, education, the most valuable is industrial applications, you can fall on mobile phones, you can also fall on robots, industrial modules, which can bring new leaps.
IDC predicts that more than 50% of the terminal processors in the Chinese market will have AI engines by 2026. We think it will bring China's Internet industry out of the trough and usher in an exciting blowout development.
The Mathematical Intelligence economy has talked a lot. In fact, our big model has added new capabilities to the digital economy in the future. I reviewed here the top 10 banks with the highest market capitalization in the world in the 1990s, mainly Japanese banks. In 2000, they were mainly IT companies in the United States. In 2010, they were energy and finance. In 2020, they returned to the dominance of the Internet. China's Ali and Tencent are also on it. By December this year, you can see them now. In addition to food and drug companies, basically still IT and IC enterprises, we say that now in the forefront is mainly digital intelligence enterprises, data has become the main factor of production.
Thank you.
Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.
Views: 0
*The comments in the above article only represent the author's personal views and do not represent the views and positions of this website. If you have more insights, please feel free to contribute and share.
Continue with the installation of the previous hadoop.First, install zookooper1. Decompress zookoope
"Every 5-10 years, there's a rare product, a really special, very unusual product that's the most un
© 2024 shulou.com SLNews company. All rights reserved.