We introduce the results of the Silicon Valley AI study tour from three aspects: the underlying technology path and computing power, industry applications, and the development trend of large models and vertical models.
Author: Minsheng Securities Lu Wei/Hu Youwen Team
In order to get a close look at the latest trends and first-hand information about the frontier development of AI in Silicon Valley, we recently made a special trip to Silicon Valley for study and investigation. We visited the headquarters of companies such as Microsoft, Google, Nvidia, and Meta, and conducted in-depth exchanges with many people in the industry, gaining a lot , I feel deeply. Many of the cutting-edge cognitions about computing power and large-scale models learned through exchanges are not the same as the understanding of the current domestic capital market, and they are even very different. These differences may contain important investment opportunities. For this reason, I would like to share with you the results of this study and inspection as soon as possible.
Our biggest feeling during this trip is that in the United States, AI is no longer a new thing that has aroused curiosity and controversy at the beginning, but has been integrated into all aspects of social production, company operations and people's lives like water and electricity. At the same time, the speed of AI technology progress is in a "Cambrian" era. Almost every week or even every day, exciting new technologies or products are born and quickly applied to actual scenarios. Therefore, American VCs , PE can be said to be almost pervasive, going deep into all aspects of AI to find investment opportunities.
Global technology giants represented by Google generally believe that this AI revolution is a major change unprecedented in human beings, and it is an unprecedented inflection point in human progress. We once pointed out that this time, AI combines the highest Internet traffic entrance with the largest imaginable space in computer history, the largest public cloud market, and the operating system that unifies all things. Even so, we may still underestimate its significance: this or It will be the greatest technological revolution in the history of human development!
We will introduce the results of the Silicon Valley AI study tour from three aspects: the underlying technology path and computing power, industry applications, and the development trend of large models and vertical models:
1 The underlying technology and computing power requirements of the large model
First, the underlying technology of the GPT large model is Google's TransFormer. Its core significance is to infer the next most likely character for output after given the previous text. Decodes a bidirectional structure. ChatGPT can support the analysis of 32,000 characters in the previous text and then infer the next word, which is already a very large order of magnitude, and ChatGPT only allows analysis and reasoning based on the previous text and does not allow analysis and reasoning by adding the latter text (decoder only). The first large-scale model Bard launched by Google can be used for analysis and reasoning (encoding and decoding two-way structure), but it was later changed to a PaLM large model similar to ChatGPT that can only use the previous reasoning. It may eventually be that this model is closer to The way we humans speak.
Second, the large model is essentially more like "alchemy" that relies on AI infrastructure. It is reasoning rather than cognition; currently, video generation is subject to high requirements for infrastructure and there are still certain limitations. Although ChatGPT's performance in QA, summary and other fields is astonishing, it is actually reasoning rather than cognition, including generating pictures, which is actually generating text in essence, but it is very difficult for AI reasoning in generating videos, because each frame of the video is It is a large number of pictures, which requires a very powerful AI infrastructure to support. Visually speaking, large-scale model training is like alchemy. The better the AI infrastructure, the better the firepower. But at the beginning, I didn't know whether the large-scale model was useful or not. There was a certain element of luck in it.
Thirdly, it is not as difficult for latecomers to catch up with large-scale models, and Chinese large-scale models will catch up with overseas countries relatively quickly. OpenAI itself is not absolutely ahead of other technology giants in technology. The core is to focus on the development of TransForemer in the direction of general artificial intelligence, while Microsoft, Google, and Meta have many profitable businesses that do not pay much attention to large models. After OpenAI was created, big companies found that AI has a future, and their own resources are better and they will definitely step up to catch up. Silicon Valley experts predict that after 6 months to 1 year, the level of large-scale models of global manufacturers will basically be the same. China's large-scale models will catch up with overseas countries relatively quickly. China itself is a very good market. Now everyone knows all the technologies of large-scale models, which are nothing more than the concentrated accumulation of resources. OpenAI has "committed" to Microsoft for doing so well, because the training itself is too expensive.
Fourth, overseas AI giants have computing power reserves of A100 chips in the order of more than 500,000 pieces. Nvidia is developing its computing power resources in the direction of cloud services, and at the same time it is also laying out its own large-scale model. At present, the average A100 size of overseas giants is estimated to be more than 500,000 pieces, and each H100 may be one to two hundred pieces, and it will be launched on a large scale in June and July. The actual advantage of Nvidia is the combination of software and hardware. Its hardware has a layer of framework tensor RT. Nvidia has an engineering design team of hundreds of people to make the framework. Nvidia not only makes hardware, but also the underlying infrastructure for Tensor RT. In the future, Nvidia is expected to form a cloud brand. At the same time, it is also laying out large models, which may have a great impact on the entire AI ecosystem.
Fifth, the market for inference chips is much larger than that for training chips, or even the sum of the training market and the cloud inference market. China has a very large market space for edge AI computing power. The application of edge computing to small devices such as the Internet of Things does not require high manufacturing processes. Now the market structure is scattered, and the market for inference chips is much larger than that of training chips, and even far larger than the sum of the training market and cloud inference market. China can use its own Advantages in the manufacturing industry, reducing the manufacturing process of the Internet of Things, and then introducing a dedicated AI reasoning chip with a small volume and low computing power to the market, this is a huge opportunity. In fact, the volume of terminal equipment is huge. For providers that can provide cloud services in the world, the number of data centers is compared with the number of massive terminal equipment. The chip demand is still very small, about 2/8 ratio.
In terms of the underlying technology and computing power requirements of large models, we believe that:
1. There is no ceiling for computing power demand. At present, the main computing power demand of large models comes from text training. In the future, from text to image to video, from training to inference, from cloud to edge, the continuous high growth of computing power demand is very deterministic.
2. The market structure of GPU chips may change. With the strong support of Microsoft and other giants, AMD’s relatively weak software ecosystem is expected to make great progress, and AMD will form a strong challenge to NVIDIA.
3. Chips are the biggest gap in the competition between China and the United States. Reaching an order of magnitude of computing power reserves between the two countries is not only a bottleneck that needs to be solved urgently, but also an investment opportunity that will be determined in the future. Especially the inference computing power on the edge side is not only an underestimated market that far exceeds the training computing power, but also gives China the opportunity to show its manufacturing advantages.
2 About AI industry application
First, the large model is suitable for industries that require a certain fault tolerance rate. ChatGPT started to make commercial paid use of plus, but it is not profitable. The core is to block some users who use it indiscriminately and make the cost too high. Large-scale model applications are currently relatively difficult in industries that require 100% accuracy, such as customer service consultation, artistic creation, meeting minutes, article writing, data analysis, etc. The commercialization of large models has already seen results on the B side, such as: Microsoft’s family bucket office, which reduces production time, improves completion, and increases repurchase rates; customer service: saves front-end customer service costs for real estate companies and medical companies. Video production: One-click generation of visla.us can only generate tools such as demo videos, so there is no need to find a studio, saving labor costs. GPT4 is only one and a half months old, and the market is still discussing how to apply it. In another six months, we can see more and clearer implementations.
Second, Microsoft's M365 products mainly focus on large-scale delivery, privacy and security. Microsoft's main goal now is how to deliver on a large scale, especially to solve some personalized AI features, and to prepare for security and privacy. M365 is the core product of Microsoft now. For the entire workflow of the enterprise, the entire collaboration platform, the entire tool, storage, and security are all under the M365 directory. Copilot is to greatly improve the production capacity of the existing product line. M365 has two different sets of computing, relying on the Azure data center for global expansion, and M365 also has its own data center inside; M365 embeds openAI in products, not using public openAI. There are technical difficulties for M365 in China: 1) Computing resources; 2) Regulations: data transparency and management of sensitive information.
We believe that in the United States, the application of AI technology has become very common, such as customer service consulting, art creation, meeting records, article writing, data analysis and many other industries. However, it should be noted that the current application of large models should be positioned as a "copilot", which requires a certain fault tolerance rate rather than deterministic decision-making. In addition, overseas large-scale model applications represented by Microsoft are still facing great difficulties in entering China. These difficulties are not only in terms of data security and compliance policy requirements, but also in the localized deployment of large-scale models and computing resources. challenge.
3 Development Trends of Large Models and Vertical Models
First, the large models of Google and Microsoft are likely to be closed source, and Meta may be the most important open source "spoiler". Google has no way out because the search will be subverted by the big model, and there is no advantage to open the big model, and AI will become an important money-making tool in the future, so there is a high probability of closing the source. Microsoft relies entirely on OpenAI, hoping that GPT will enable efficient office tools such as MS365 Copilot and the Bing search engine, and Microsoft will most likely not open source AI. The most important business of Meta is social networking, and AI can be used as a chat assistant. Meta's idea is to make a large model and then open source it, becoming a "spoiler" in the large model. In comparison, Meta's large model has 175 billion parameters, and it is estimated that GPT4 has about 500 billion parameters. Meta has open sourced a large model with more than 65 billion parameters, and the estimated accuracy is about 20% lower than that of ChatGPT. Many companies and studies use Meta's open source model for fine-tuning, and the effect is similar to that of GPT on the basis of small model parameters. The significance of open source is that it can mobilize millions of engineers around the world to participate in fine-tuning.
Second, it is a trend for large models to move to the mobile terminal. In the open source ecology of large models in the future, large companies will make large models, and small companies will make fine-tuning. The large models will also be simplified to each mobile terminal, such as changing the original 32-bit floating-point calculation to INT8, etc., to improve the calculation speed. Large language models will have a good ecosystem in open source. Large language models are like water and electricity, and open source ecosystems can be used in some subdivided areas. Some smart people in the open source community can distill the model to a very small size, such as changing the 36-bit floating-point calculation to INT4, which can reduce the size by ten times, so that it can be installed on computers and mobile phones. There may be many creative models in the future. application developed. In the future, iOS or Android may have a built-in large model, and in the future, all mobile applications will be charged to Apple once.
Third, ROI should be considered at the core of increasing the number of parameters in the continuous development of large models. From the perspective of scientific research, of course, the more parameters the better, but from the perspective of commercial use, each additional parameter will increase the cost, including collection costs and training costs. ChatGPT 3.0 uses 1750B parameters, and there is a GBT-like model on GitHub that only uses 70B parameters to achieve 90% GPT effect. From the commercial application level, it is necessary to find the parameters with the highest ROI.
Fourth, the large model will eventually take over some vertical industries whose data can be obtained through the Internet, and may not be able to cover some vertical field models whose data cannot be obtained. Now Google is doing something to allow AI to learn the content of the Internet in real time like a human. In areas where data cannot be obtained offline, there may be a form of interaction between an online large model and a local model, but this involves a relatively difficult one. Coupling problem.
We believe that:
1. China may be the biggest beneficiary of the open source model represented by Meta.
2. We should maintain confidence in the progress of domestic large-scale models catching up with the world's leading level, and catch up on the basis of the given technical route direction and open-source large-scale models, which actually saves the cost of trial and error from scratch. Especially for vertical industry leaders who do not have high requirements for the versatility of large models, they can quickly build vertical large models with the help of open source large models, and accelerate the application of vertical fields.
3. It is an inevitable trend for large models to be deployed on the edge side and on the mobile side, especially after Google recently released large models on the mobile side and the ChatGPT App on Apple mobile phones was officially launched. This trend has gradually been recognized by the market. The long-awaited unified operating system for standardized various AIoT terminals.
Special statement: In any case, the relevant content should not be regarded as investment advice
analyst commitment
The signed analyst of this report has the securities investment consulting qualification granted by the China Securities Industry Association and is registered as a registered analyst. report, and is responsible for the content and opinions of this report. This report clearly and accurately reflects the researchers' research opinions, and the conclusions are not inspired or influenced by any third party. receive compensation of any kind.
Recommended reading:
4D Interpretation Meta Transformation Metaverse
Metaverse is a "value creation" economy
< END >
