Data-driven AI technology: how to help enterprises rapidly transform and upgrade

The open source China community team made its first live broadcast, telling the story behind the open source China community in the name of sharing."

In the distant prehistoric era, our ancestors started a dialogue with nature through slash-and-burn farming, and also announced the birth of tools as an important cornerstone of the progress of human civilization.

Tools, this seemingly simple but far-reaching existence, have been closely linked to the destiny of mankind since their birth. From the initial stone tools and wooden sticks to the later bronze and iron tools, and now to today's high-tech products, every evolution of tools marks a leap in human productivity and promotes the progress of social civilization. They are not only an extension of human hands, but also the crystallization of wisdom and a weapon to conquer nature and transform the world.

AI tools: a powerful tool for rapid enterprise development

Looking back at the present moment, AI has become a bright star in the technology world. IT, Internet and other fields have embraced it. AI-based applications are springing up like mushrooms after a rain. People are increasingly lamenting: Nowadays, APPs are becoming more and more popular. The more "smart" you become!

However, for many traditional industries, integrating AI into daily business is not easy. The technical threshold is high and scene integration still needs to be deeply explored. Therefore, they are more eager to obtain AI tools that are easy to use, efficient and practical.

Chen Xiaojian, General Manager of Amazon Cloud Technology Greater China Product Department

"Amazon Cloud Technology's three core capabilities in building a data base cover important scenarios from basic model training to generative AI application construction, which can help enterprises easily deal with massive multi-modal data and improve basic model capabilities. Currently, Amazon Cloud Technology is developing Helping enterprises of all sizes in various industries to build a strong data base, while ensuring the security of user business and data, assigning the unique value of data to basic models and generative AI applications, accelerating enterprise business growth." Talking about AI. Combining it with enterprise applications, Chen Xiaojian, general manager of Amazon Cloud Technology Greater China Product Department, gave this interpretation.

Obviously, in addition to data foundation, enterprises also need to choose appropriate AI tools based on their own application scenarios in order to leverage the power of AI to promote business development. Perplexity is an example of a company that creates unique value by blending traditional search, customer data, and the inference and text transformation capabilities of large language models. The company is building the world's first conversational answer engine. Since its launch in December 2022, its website and mobile applications have quickly won the favor of users, reaching 10 million monthly active users, attracting 53 million in November alone Views. This rapid growth is unmatched by traditional marketing methods.

Three ways to deal with different application scenarios

Chen Xiaojian said: When exploring how to integrate an enterprise's own data into generative AI applications, we discovered three methods: retrieval-augmented generation (RAG), fine-tuning, and continuous pre-training, which can combine data with large language models. to improve business results.

In generative AI, high-quality output often relies on a large amount of contextual information. Enterprises can combine their own knowledge bases (such as databases or other knowledge documents) with generative AI to provide auxiliary capabilities for large language models through upward search and other methods. This approach is relatively simple and many businesses are already using it to build applications. This is Retrieval Augmented Generation (RAG).

Fine-tuning techniques involve additional training of the model using data associated with a specific target task, with the goal of improving the model's performance on the specific task. The difficulty of fine-tuning lies between pre-training and RAG (retrieval augmented generation). It is suitable for a variety of scenarios, such as character understanding, input content analysis, and output format control.

Continuous pre-training has a high threshold and requires a large amount of data. Enterprises need to continuously input data generated in daily business into large models for training to adapt to business changes. Instead of building a training environment from scratch, continuous pre-training is based on an existing large model.

Amazon Cloud Technology’s Amazon Bedrock product has implemented three key capabilities that many customers are using to train customized large models in their business environments. Focusing on models means focusing on business results, and a solid data foundation is the key to success. Therefore, Amazon Cloud Technology has an important point of view: no data and no models.

The cornerstone of AI applications: data storage

在AI时代，可以看到存储解决方案不仅需要承载海量数据，还必须提供足够的性能，并要有可控的成本。由于多模态模型的流行，数据类型在规模和形态上存在显著差异，这要求我们拥有强大的数据存储能力。

Chen Xiaojian said that Amazon S3 is the earliest data storage cloud service launched by Amazon Cloud Technology. It has evolved into a platform that fully meets the data storage requirements for fine-tuning or pre-training base models. Amazon S3 holds over 200 trillion objects and handles over 100 million requests per second. It also provides fine-grained control, compliance auditing capabilities and lifecycle management capabilities to ensure data security and legal use. Amazon S3 is also an ideal choice for building data lakes. There are more than 200,000 data lake applications on Amazon Cloud Technology.

Amazon S3可支持高效、经济地大规模数据分析，适用于人工智能、机器学习和高性能计算等多种应用场景。在生成式AI时代，对数据存储和处理性能的需求日益增长。为了满足这一需求，亚马逊云科技还推出了Amazon S3 Express One Zone，这是一项新的服务，能够实现低于10毫秒级别的快速访问，许多客户已经通过这项服务结合他们的业务实现了显著的性能提升。

In the AI era, serverless architecture helps enterprises grow rapidly

In modern data processing environments, relational databases are only one option for vector retrieval capabilities. With the widespread application of search functions, various database types such as relational, key-value, graph databases and document databases are playing an important role in their respective fields. However, when it comes to vector retrieval, specifically introducing an entirely new vector database can come with learning costs, the cost of configuring new resources, and the complexity of data migration.

In current observations, many customers prefer integrating vector retrieval capabilities into their existing databases rather than introducing entirely new database systems. The benefit of this is that additional learning costs, migration costs and possible licensing fees are avoided. At the same time, centralized storage and management of data helps shorten response time and improve performance.

Especially in the era of GenAI (generative artificial intelligence), rapid launch and market capture have become the primary goals of many companies. Therefore, it becomes particularly important to provide vector retrieval capabilities for various databases. This not only meets customers' performance needs, but also ensures unified management and efficient retrieval of data.

In addition, with the growing need for rapid development and deployment, serverless (serverless) architecture solutions are favored for their flexibility and cost-effectiveness. For companies that do not have dedicated personnel for operation and maintenance or DBA work, the Serverless solution does not require predicting future performance needs or performing tedious operation and maintenance operations. During peak business periods, it can automatically expand to meet demand, and when business is idle, resources can be automatically recycled to save costs.

Therefore, for customers at the current stage, providing vector retrieval capabilities and serverless capabilities for various databases is the key to meeting their rapid development and deployment needs. This can not only improve data processing efficiency, but also reduce operation and maintenance costs and complexity, giving companies an edge in the fiercely competitive market.

In terms of AI application implementation, Amazon Music uses advanced technology to analyze the characteristics of users and songs, and converts this information into vectors to improve the accuracy of music recommendations. By using Amazon OpenSearch, Amazon Music successfully transformed 100 million songs into vectors and indexed them to provide real-time music recommendation services to users around the world.

Currently, Amazon Music maintains 1.05 billion vectors in Amazon OpenSearch and has the ability to handle up to 7,100 queries per second, effectively supporting the operation of its recommendation system.

The construction of generative AI is not easy. It is more like a flywheel structure and requires a positive cycle to promote its development. To achieve this, enterprises need to leverage multiple cloud services to build a solid data foundation. In this way, enterprises can efficiently and securely combine massive data with underlying models to create generative AI applications with unique value that meet the needs of end customers and generate more data.

As these applications come into use, they generate new data, which in turn further improves the model's accuracy. Through continuous fine-tuning or pre-training, models can become smarter and more industry-professional, thereby providing users with a better experience. This continuous loop of positive feedback mechanism will bring continuous power to the enterprise and promote the continuous success of its business.

Data-driven AI technology: how to help enterprises rapidly transform and upgrade

Guess you like