From the end of 2022 to the first half of 2023, many people have witnessed the global popularity of ChatGPT and the explosion of generative AI and large language models.
So for the majority of users, what kind of opportunities does generative AI mean? What role does data play in generative AI applications? How to unlock the value of generative AI? How should we start the journey of generative AI?
At the 2023 Amazon Cloud Technology China Summit held recently, Matt Wood, vice president of global products of Amazon Cloud Technology, gave detailed answers to the above questions.
Unlocking the Value of Generative AI
"Amazon Cloud Technology firmly believes that generative AI is reshaping all kinds of industries. It can break through the limits that could not be broken before. Its typical scenarios include creative output such as writing, design, coding, and modeling; including search, summary, sorting and other functional enhancements; including new interactive experiences that can generate new knowledge and new ideas, and provide open knowledge through chatbots; including decision support for automatic execution of complex tasks under human supervision."
When it comes to how to unlock the value of generative AI, Matt Wood said that there are four main steps:
1. Provide access to first-class basic models: Amazon Bedrock service of Amazon cloud technology can provide access to first-class basic models, allowing users to access basic models from AI21 Labs, Anthropic, Stability AI and Amazon through API, so it is the easiest way for customers to use basic models to build and expand generative AI applications, helping all developers lower the threshold of use. On Amazon Bedrock, users can access a series of powerful basic models from text to images through scalable, reliable and secure Amazon cloud technology hosting services.
2. Provide a safe and private environment to customize models: Amazon Titan, a subsidiary of Amazon Cloud Technology, is a series of different model libraries that can implement text summarization, search result embedding, harmful content deletion, etc. Users can optimize and fine-tune these models in a very safe and private manner.
3. Provide low-cost and low-latency access through custom chips: From Nitro and Graviton to machine learning reasoning chip Inferentia and machine learning training chip Trainium, Amazon Cloud Technology’s low-cost, low-latency self-developed custom chips are showing more and more obvious advantages in the field of generative AI.
4. Search for opportunities to improve user experience: Based on machine learning technology, Amazon Cloud Technology provides developers with Amazon CodeWhisperer code generation service, which supports 15 different programming languages including Java, JavaScript, and Python, and can be used by individual users for free. In tests, participants who used CodeWhisperer completed tasks an average of 57 percent faster and had a 27 percent higher success rate than participants who did not use CodeWhisperer.
"Amazon Cloud Technology has a long history in providing customers with a wide range of machine learning capabilities. From continuous innovation in machine learning, customized underlying chips, out-of-the-box AI services to leveraging Amazon Bedrock usage models, if you put all these together, you will find that there has never been such a simple and low-cost way to build code with machine learning." Matt Wood said.
Amazon Cloud Technology's Cloud Native Data Strategy
Matt Wood pointed out that data is not only the starting point of generative AI, but also will promote the widespread application of generative AI. To this end, Amazon Cloud Technology has also launched its own cloud-native data strategy:
1. A comprehensive set of tools to meet current and future needs: Amazon Cloud Technology has launched 15 purpose-built cloud-hosted database services to provide perfectly suited data services for various user application scenarios; Amazon Cloud Technology’s analysis services have been fully serverless, including interactive query service Amazon Athena, big data processing service Amazon Managed Streaming for Apache Kafka (Amazon MSK), real-time analysis service Amazon Kinesis, data warehouse service Amazon Redshift, data integration service Amazon Glue, business Intelligence service Amazon QuickSight and operational analysis service Amazon OpenSearch Service.
2. Easy integration and connection of all data: Amazon Cloud Technology put forward the vision of Zero-ETL, which is committed to realizing seamless data conversion and calling, without users needing to write any code. For example, the recently launched Amazon Aurora service can integrate Zero-ETL with Amazon Redshift, allowing Amazon Redshift to perform near real-time analysis and machine learning on PB-level transaction data from Aurora. Transactional data is available in Amazon Redshift within seconds of being written to Aurora, so users don't have to build and maintain complex data pipelines to perform extract, transform, and load (ETL) operations.
3. Build end-to-end data governance: through data governance, data circulation is accelerated and guaranteed. Amazon Cloud Technology launched a new data management service Amazon DataZone in 2022, which allows customers to catalog, discover, share and manage data stored on Amazon Cloud Technology, customers' local and third-party sources faster and easier. With Amazon DataZone, administrators and data asset managers can use granular control tools to manage and govern data access permissions, ensuring that data access occurs with the correct permissions and in the correct context.
Tips for Starting Your Generative AI Journey
When it comes to how users should use data to start their generative AI journey, Matt Wood gave five suggestions:
1. Build based on your existing data strategy;
2. Enabling broader and safer generative AI experimentation within organizations;
3. Customized models for demand scenarios;
4. Join hands with Amazon cloud technology to explore infinite possibilities;
5. Select the scene, set off immediately, and build the future.
"Technical capabilities often follow the S-curve. You never know where you are on the S-curve, unless you look back afterwards. I hope that when we meet again next year, we can review what new points we have at the head of the S-curve, and what prospects we have in the future. We hope to have a gradual push forward on this curve. We believe that when we look back next year, we can see that we have reached the apex of the S-curve, reached the point of rapid and explosive development, and many innovative inventions and applications have emerged." Matt Wood, "science fiction writer William Gibson I once said that 'the future is close at hand, but it is only appearing from time to time', Amazon Cloud Technology always hopes to provide everyone with promising and exciting technologies. In the era of cloud computing, we provide quick access to cloud applications through APIs. This vision will not change because of the emergence of generative AI. We hope to put this technology in the hands of every builder and every business user."
"This Amazon Cloud Technology China Summit demonstrates Amazon Cloud Technology's continuous deepening in the field of cloud and AI. Especially in the field of AIGC, which has recently attracted heated discussions, Amazon Cloud Technology has placed more emphasis on empowering customers and partners with its technical capabilities and best practices for AIGC business development. Through more flexible, agile, and low-threshold products (such as Amazon Bedrock and Amazon CodeWhisperer), it will accelerate the popularization of AIGC technology and explore value in more industry scenarios. It is committed to becoming the 'behind the scenes' and 'digital pedestal' in the AIGC world." Ai Wang Chengfeng, research director of Rui Consulting, said.
The "Hundred Models War" in the field of generative AI
Since ChatGPT became popular overnight, a variety of generative AI tools and large models have appeared on the market, known as AIGC's "Hundred Models War", which also makes many users wonder how to choose. What views and suggestions does Amazon Cloud Technology have on this?
"With the emergence of ChatGPT, many people are deeply encouraged, but we must also seriously consider what kind of generative AI tools we need, what kind of system we need to build, what kind of resources we need, especially what kind of large model, in order to realize our ideas." Matt Wood said, "The goal of Amazon cloud technology is to help customers build their own large models in the simplest possible way. Whether it is a startup, a small or medium-sized enterprise or a large enterprise, these generative AI tools can be used. Our developers can provide easy-to-use tools in incredible ways. Generative AI tools. Whether it’s text or images, users only need to understand a simple API and select a suitable model to output what they want. In this process, what distinguishes Amazon Cloud Technology is that the generative AI tools we provide must be available everywhere and be implemented at a very low cost, delivering services to users with the lowest possible delay while ensuring the optimization of operation and maintenance operations. Therefore, we chose the method of customizing chips, which can provide more powerful performance and reduce costs and delays as much as possible.”
Matt Wood pointed out that from the perspective of Amazon Cloud Technology, the needs of users in different industries vary widely, so there is no "one trick" universal big language model that can be applied to multiple application environments. For this reason, Amazon Cloud Technology is actively using large language models from third-party partners such as Anthropic, AI21Labs, and Stability AI on the basis of training Amazon Titan and other self-developed large language models, so that users can use their own data according to their own needs and customize corresponding large language models in various ways.
"It is thanks to our R&D and innovation in the field of generative AI over the past period of time that we have achieved low cost and low latency, enabling customers to build and use their own models through custom chips. At present, in the entire industry, Amazon Cloud Technology has achieved the fastest, lowest cost, and easiest way among all suppliers to provide large generative AI models." Matt Wood said, "At the same time, we are constantly launching new services, such as using the generative AI product CodeWhisperer to help software development People have greatly improved their productivity, enabling them to write software and programs faster, with higher quality, and in a safer way. This has also made many customers very excited and satisfied."
