Recently, the Ali version of ChatGPT "Tongyi Qianwen" officially announced that it will open to the outside world for enterprise invitation testing, and many people have already obtained the invitation code to enter the test. So, looking back at the current news about the application of large models, what stage has the current development of this field reached? Let's take a look at the author's analysis and opinions.
The release time of Alibaba’s large model is getting closer and closer, and the voices in the market are also mixed. Recently, Zhang Yong just announced the most important organizational change decision in the 24 years since the establishment of Alibaba. Xiaoyaozi is directly responsible for the cloud intelligence group to which the large model belongs. , the matter becomes more subtle.
"Xin Mou" also got the test invitation code of Ali's large model for the first time. The most intuitive feeling after experiencing it is that it is remarkable and the response speed is not bad, but whether it is Baidu or Ali, the large model is still at the transformer level. , did not emerge an excellent comprehension ability for the text.
Generally speaking, Tongyi Qianwen and Wenxin Yiyan are equally matched in terms of ability, but this is just the beginning. Compared with the text question and answer ability, we are actually more curious about the next actions of "Tongyi Qianwen".
Because at the industry level, due to the unknown rhythm of ByteDance (the leader of the original Ali M6 large model is now joining Byte AI Lab), the focus of the first echelon falls more on Tencent, Baidu, Ali and Huawei. One is that these companies have the background and talent and technical reserves in this area, and the other is that from the perspective of business, these large factories all need the support of large models.
Now Baidu has bravely taken the first step and released the preliminary version of Wenxin Yiyan. However, due to Baidu's aggressive branding style at that time, the outside world has greater expectations for Ali's large model, but according to " According to the information disclosed in Hezong Investment Research, Ali’s large model is still in the stage of catching up with GPT 3.5, and there is still a long way to go before GPT 4.
There are three key factors causing this "stuck neck": one is the amount of data, which accounts for about 30% of the weight; the other is the innovation of the large model structure. The ratio is around 40%-50%; the final 30% weight is engineering implementation, and there is no good reference in this area in China.
Reminiscent of Open AI and the evolution logic of GPT behind it, everyone has begun to realize that this AGI competition is not just a competition of financial and human resources, but from theory, to engineering, to products, and then to business. , A comprehensive subversion in the humanities, back to the Ali large model, perhaps we should focus more on the application of the large model and Ali's existing business modules.
Zhang Yaqin, dean of the Intelligent Industry Research Institute of Tsinghua University and academician of the Chinese Academy of Engineering, talked about this topic recently. He believes that the generative AI model of the GPT series can be regarded as an AI operating system composed of large models, which is similar to that on the PC back then. Windows, and mobile Android and IOS basically have similar meanings.
The birth of a new operating system often means that the underlying hardware and the above applications will be refactored to form a brand new ecosystem. He even asserted that if the ecological value of the PC Internet is 1X and the ecological value of the mobile Internet is 10X, then the AI ecology is at least 100X.
Zhang Yaqin’s judgment largely comes from the combination of Microsoft and Open AI. With the blessing of GPT 4, Microsoft’s once declining business sectors have become imaginative, including the search business that was beaten steadily before. Bing.
This also explains why major manufacturers including Baidu, Ali, and Huawei are all betting on large models. To put it bluntly, chatGPT is a hook. If you still focus on how large models will subvert the search scene, then I only It can be said that you not only don't understand the big model, but also don't understand the business gameplay. This is the real reason why many big manufacturers choose to end, even though it seems vague now.
To be precise, this is fundamentally different from the “false outlet” that was ripened by capital a few years ago. LLM, including AGI, is a real outlet that has been verified by Microsoft, and was even warned by Elon Musk at one point—artificial Intelligence is evolving in a direction beyond human control.
Obviously, Musk’s concerns are more on the ethical and moral level. This is also the most awe-inspiring part of the science of artificial intelligence. Once it is manipulated by people and companies with ulterior motives, it is very likely to bring about relatively large. disaster. But if we look at LLM and AGI under norms and legal standards, it is obviously a very good tool, like a storage box for human intelligence, which outputs capabilities through large models.
Combined with the current actions of Microsoft, the large model has a good fit in the fields of cloud computing, text office, and search. It is not difficult to speculate that the direction of the domestic large-scale model will also follow the gourd, and the pilot will start from these aspects. Coincidentally, these businesses are available in Ali, such as Alibaba Cloud in the cloud, DingTalk in the text office, and DingTalk in the search. quarks.
More importantly, Ali also has a hole card that Microsoft does not have - e-commerce . E-commerce is the foundation of Alibaba, but just like Amazon is to Bezos, in the face of the rise of various forces such as live broadcast e-commerce and interest e-commerce, the site has been impacted. This has been difficult for Alibaba to say over the years. heart disease".
Therefore, for Ali, the large model has to be done, and it must be done. Even if the path of income increase is vague, it may be a very good intelligent tool, and it is far more imaginative than the data center built in the past few years. Some, the flexibility of the front end can actually allow Ali to find new growth possibilities.
Regarding LLM and AGI, everyone’s thinking is still relatively limited. For example, for the question of “what will the big model subvert?”, when we give the answer, we may be more inclined to answer “what the big model can’t subvert”. The reasons for this situation include not only the level of public awareness, but also the status quo of industry development.
This is the same as watching sci-fi movies. As early as 10 years ago, we could see various intelligent scenes in American blockbusters, but in fact, even if it has developed to this day, we are still a long way from the so-called AGI. There is a long way to go, but what is certain is that after the big companies headed by Baidu and Ali have left the market, the real wave of artificial intelligence has begun.
In the 10-year wave of mobile Internet, many scenes in life have been subverted, including basic necessities of life, without exception. But if you look deeper, the root cause of this change is actually driven by cost factors on the supply and demand side. There will always be more cost-effective solutions that will replace the old ones. This is the new law of the digital economy era.
So if we look at the supply and demand perspective of the large model, we will find that this market is far from being well developed. Just like before Tesla came out, no one would have thought that Toyota Motor, the global hegemony of the year, would be challenged so much. According to this logic, all the judgments on the market now are not credible, or not fully credible.
Perhaps in another year, as these major manufacturers gradually apply large-scale models to make them easy to use and affordable, our understanding of large-scale models may be more objective.
Finally, let’s talk about the topic of unemployment that everyone is most concerned about. This issue is obviously too much to worry about.
Many times, we tend to overestimate artificial intelligence and underestimate our own creativity.