DoNews reported on April 8 that generative AI is sweeping thousands of industries, and to get the ticket of generative AI, "going to the cloud" has become the first choice of most enterprises.
With the rise of generative AI, Amazon has been attacking frequently since 2023: after announcing a $4 billion investment in AI startup Anthropic last year, it has recently continued to invest an additional $2.75 billion, completing a $4 billion investment in Anthropic.
As a leader in the global cloud vendors, to what extent has Amazon Web Services developed generative AI? What will be the key step of generative AI in the future? On April 2, a media conference with the theme of "Amazon Web Services Joins Hands with Anthropic to Promote Generative AI Innovation" was held in Beijing, focusing on the Claude 3 series of large models released a month ago, and answering relevant questions to the media one by one.
01. "There is no one model that can be applied to all business scenarios"
It has been more than a year since the launch of generative AI, and Amazon Web Services has cooperated with many customers in various aspects. They believe that the scenarios of various industries are ever-changing, and although the technical capabilities of large models are strong, "no one model can be applied to all business scenarios".
According to Chen Xiaojian, general manager of the product department of Amazon Web Services Greater China, enterprises still need to add a lot of auxiliary capabilities from model capabilities to real operation and production.
"Although large models are very important, large models alone are far from enough for your production, enterprises need a series of peripheral capabilities to use large models correctly, reasonably, safely, and efficiently, which is where the value provided by a series of Amazon Web Services products lies. ”
In this context, Amazon Web Services provides customers with a three-tier architecture in the field of generative AI, through which different customers can choose different tiers of products to support their business according to their needs.
The first layer is the generative AI cloud infrastructure.
Amazon Web Services provides customers with basic computing power, including NVIDIA's latest G200 chip. In addition, Amazon Web Services has also invested a lot of effort in the development of self-developed chips, including Amazon Trainium, a chip used for training, which is now in its second generation, and Amazon Inferentia, an inference chip, has also entered its second generation. In addition to this, Amazon Sagemaker, a platform for training and inference, is included.
The second layer is the model.
That is, some new products that have come out with generative AI, such as Amazon Bedrock, Amazon Bedrock supports multiple technology models through a model platform, which is a capability provided by Amazon Web Services.
The third layer is the application layer.
Amazon Web Services launched Amazon Q, a generative AI assistant, and combined with Amazon Connect, a business intelligence service, and Amazon Quicksight, a platform for programming developers, to support the business needs of different customers.
02, Claude 3, the new "volume king" of large models
It is understood that Amazon Bedrock currently provides a variety of leading basic models for customers to choose from, including well-known open-source models, such as Stable Diffusion XL, Llama, Mistral 7B and Mixtral 8*7B, as well as non-open-source models such as Anthropic Claude 3, AI21labs Jurassic, Cohere Command, Amazon Titan and so on.
The Anthropic Claude 3 sets new performance benchmarks for a wide range of cognitive tasks.
The Claude 3 is available in three versions, namely Haiku, Sonnet, and Opus, which are medium, large, and extra-large cups, and customers can choose the most suitable combination of intelligence, speed, and price according to their business needs.
- Claude 3 Haiku, with an almost instant response and the most compact
- Claude 3 Sonnet, the ideal balance between skill and speed
- Claude 3 Opus, the smartest model designed to handle highly complex tasks
At the sharing session, Amazon Web Services demonstrated Claude 3's ability to deal with math problems, programming exercises, and scientific reasoning.
Image interpretation is a problem that will be encountered in a variety of application scenarios, Claude3 can recognize the text of an image and answer the input image. Claude3 is trained to understand images such as pictures, charts, graphs, and OCR scans, and it is faster than any other multimodal model in the industry.
Affected by data, model structure, and training algorithms, hallucinations are unavoidable. Claude 3 is able to reduce hallucinations. Claude 3 is significantly more accurate when dealing with challenging open questions (100Q Hard) and reduces incorrect answers.
Amazon Web Services also showcased some Claude 3 application scenarios, including content continuation, code assistance, e-commerce product description writing, and long text knowledge recall summary.
For example, the long text knowledge recall summary can accurately answer the price of the service in different regions and extract more complete information according to the relevant service documents provided, and the code assistance can provide detailed code steps to help programmers correctly modify the configuration in the service management platform Nacos.
It is understood that Claude 3 is now multimodal capable - Claude 3 can receive image-based inputs with roughly the same capabilities as other cutting-edge models, and the latency is lower than other multimodal models (especially Claude 3 Haiku), including:
Trained for common enterprise use cases: Trained to understand pictures, charts, graphs, technical diagrams, and optical character recognition (OCR).
Speed is better than other multimodal models: The evaluation showed that the Claude 3 model was comparable to the frontier model in terms of image input capabilities, and that the Claude 3 Haiku was faster than all frontier models with comparable capabilities.
Excellent performance in use cases that require speed and intelligence: The Claude 3 model combines low latency and powerful features to excel in enterprise use cases that need to process large volumes of images, charts, reports, and other visual assets.
03. How to take the key step of generative AI?
The current cloud computing vendor, Alibaba Cloud, asked Lao Luo Live to sell cloud and JD.com to compare prices in real time...... The various price wars reflect the white-hot extent of the market. From the current point of view, the era of "one trick is eaten all over the world" has passed, and for large manufacturers, if they only do cloud, or only make large models, and only make chips, there will be shortcomings.
In the process of expanding from cloud infrastructure to chips and large models, a new battlefield has opened, and cloud vendors have their own paths on the journey of generative AI.
Returning to the perspective of enterprise needs, enterprises have at least several core requirements for using basic large models, such as data security and compliance, and easy-to-use AI platforms and toolsets. Amazon Bedrock is the first choice for enterprises to build and apply generative AI by providing access to the world's leading foundational models, as well as convenient tools such as knowledge bases, brokers, Guardrails, and more, while ensuring data privacy and security.
In addition, Amazon Web Services has a wealth of professional technical support resources, including architects, product experts, artificial intelligence labs, data labs, rapid prototyping teams, and professional services teams, to help customers solve the engineering challenges of the last three kilometers of applying generative AI.
What is the most critical step in the next development of the generative AI field? Chen Xiaojian explained the three-tier architecture of generative AI as an example.
First of all, from the bottom level, chip performance still lags behind demand. Although the development of semiconductor chips has been very fast, the expansion of the parameter scale of the model itself is actually far beyond the capabilities of the chip. "As the scale increases, the future of models will become more and more complex. As a basic service provider, we still need to pay attention to how to achieve this goal, how to match the underlying capabilities, business complexity, and large model complexity, so that hardware development can catch up with software development. ”
Secondly, from the point of view of the model itself, there is still a lot of room for development of its capabilities. "The model we see today may represent a Ph.D. student, but can it be done better, such as reaching the level of a professor or an academician? Many generative AI vendors, including Amazon Web Services, need to work consistently on model capabilities in the future. ”
In terms of the combination of the top layer and various industries, Chen Xiaojian said that the combination of Amazon Q and BI (business intelligence) service Amazon QuickSight provided by Amazon Web Services, and the combination of Amazon Connect intelligent customer service, are similar to SaaS solutions. "We need to think about how large models can be used in a more accessible way and at a lower cost to provide better model capabilities for applications in all walks of life in human society. ”
Chen added, "I think there's a saying that says it well, the iPhone era of generative AI has arrived. Today's demo gives us an idea of how much generative AI can accomplish that was previously impossible. But to really do this, I believe that not only Amazon Web Services, but the entire industry needs a lot of work to do. ”