After the basic model is built,Tiger'sTroveTrekThe key to the construction of generative AI comes to the data level.
April 30thTiger'sTroveTrekChen Xiaojian, general manager of Amazon Cloud Technology Greater China Product Department, emphasized the key role of data in the generative AI era at the media communication meeting of Amazon Cloud Technology: "No data without model-the data base in the generative AI era".
He said that data is at the core of the generative AI era, and if enterprises want to succeed in the generative AI era, they must start with data and use their own data to build AI applications with commercial value.
Chen Xiaojian believes that the data capabilities of enterprises need to be constructed in the following three aspects: the data processing capabilities required for model fine-tuning and pre-training, the ability to quickly combine proprietary data with models to produce unique value, and the ability to effectively process new data to promote the sustained and rapid development of generative AI applications.
When asked whether Amazon Cloud Technology does not have the advantage in artificial intelligence reasoning and training costs as the cloud market enters the AIGC era, Amazon Cloud Technology said that it still cares about what customers want in the end. In other words, Amazon Cloud Technology hopes that no matter how the cost of services changes in the AIGC era, its business model is still focused on providing tools and services for cloud infrastructure, model layer and application layer.
Why data processing is important
The importance of a large amount of high-quality data to generative AI has become an industry consensus.
Amazon Cloud Technology said this time that if every company has access to the same basic model, then all companies are at the same starting line, while companies that can use their own data to build generative AI applications with real business value win at the starting line.
It can be said that one of the limitations of the generative AI basic model is that it is unable to own the proprietary data of the enterprise in time. If you want the model to serve the development of the enterprise, then accelerating the combination of data and model through technical means has become one of the key capabilities of the enterprise data base.
Chen Xiaojian further said that successful enterprises need to understand business and users' generative AI applications, and these applications need to be built from data. He cites the example of Perplexity, a US-based artificial intelligence start-up, which has achieved rapid growth and user attraction by combining traditional search, customer data and large language models.
As a result, the company is a "celebrity" in the field of artificial intelligence. It is reported that Perplexity is working on at least 2.Tiger'sTroveTrekA new round of financing of $.50 billion could be valued at between $2.5 billion and $3 billion. The company has just made two large investments in the past four months, and its valuation has made a leap: in January this year, Perplexity made 5%Tiger'sTroveTrekA valuation of $.40 billion raised nearly $74 million; in early March, Perplexity raised about $63 million at a valuation of $1 billion.
At present, the ways of using enterprise's own data to differentiate generative AI applications and customize basic models through data are mainly divided into three categories: retrieval enhanced generation (RAG), fine-tuning and continuous pre-training. These three methods have different applicability and data requirements in different application scenarios.
Countermeasures of Amazon Cloud Technology
In this regard, Amazon Cloud Technology has emphasized the three core competencies it has built on the data base: the data processing capabilities required for model fine-tuning and pre-training, the ability to quickly combine proprietary data with models to generate unique value, and the ability to effectively process new data to facilitate the sustained and rapid development of generative AI applications.
In terms of data storage, the Amazon S3 service provided by Amazon Cloud Technology can meet the stringent data storage requirements of users in fine-tuning and pre-training the basic model. At the same time, the sub-millisecond latency and high throughput performance of Amazon FSx for Lustre file storage service will further accelerate the speed of model optimization.
In terms of data cleaning and governance, Amazon Cloud Technology helps enterprises to efficiently complete data cleaning, deduplication and word segmentation through services such as Amazon EMR Serverless and Amazon Glue, enabling enterprises to focus on AI business innovation.
In addition, Amazon Cloud Technology highlights its innovations in vector search and serverless architecture. Among them, retrieval enhanced generation (Retrieval-Augmented Generation,RAG) technology is generally considered to be one of the main ways to realize the combination of data and model. By converting the data into vectors and storing them in the vector database, RAG transforms the semantic relevance into the mathematical distance between vectors, so as to realize the relevance calculation of content.
The combination of vector search and data storage can bring many benefits, including more efficient and accurate retrieval capabilities, processing and indexing large-scale data, and so on. At present, combining the advantages of vector search and data storage, we can build a powerful information retrieval system to meet the needs of modern applications for speed, accuracy, reliability and intelligence.
But it also brings some cost problems, such as the need for more storage space. Chen Xiaojian told the Daily Economic News that storage does increase costs, but it can achieve better results in the retrieval of the entire data content.
Amazon Cloud Technology also said it has added vector search capabilities to eight data stores. Customers can also reduce model invocation costs and response latency for spanning AI applications through Amazon Memory DB in-memory databases, and accelerate innovation with and serverless technologies. In Amazon Cloud's emphasis on the data base in the generative AI era, we can see the importance of data processing as well as the challenges and opportunities faced by enterprises in this field. For Amazon Cloud Technology, providing more comprehensive services for enterprises in the AIGC era is both an opportunity and a challenge.