Generative AI technology is undoubtedly one of the biggest imagination of the current era. Capital, entrepreneurs, ordinary people are pouring into the generative AI to find out: “100 mode war” played overnight, the scale of financing a record high, a variety of consumer product concepts continue to emerge …… According to the Bloomberg Intelligence report According to Bloomberg Intelligence’s report, the size of the generative AI market in 2022 was only $40 billion, and it is expected that this figure will exceed $1.3 trillion by 2032, with a compound annual growth rate of up to 42% over the next 10 years. However, looking lively on the surface, but is the popularization and transformation of generative AI technology really as high as we think?
After experiencing explosive growth, almost all visits to generative AI chat products have declined to varying degrees since June. The latest user survey shows that more than 80-90% of respondents said that they will not use chat tools such as ChatGPT and Bard at all in the next six months. From the consumer side, it seems that people are now treating generative AI products more as a toy to keep up with fashion trends than as a tool to be used consistently. On the enterprise side, this is even more pronounced. Once people switch to work mode, generative AI tools rarely appear in everyone’s workflow, and are even explicitly banned or restricted by many large companies and others.
More than six months is not a short period of time for a commercial conversion of a more mature technology. However, at present, the focus of the generative AI mania still seems to be stuck in the big models and product concepts, people expect to see a prosperous ecosystem and the transformative impact on the economy and society has not yet arrived. So what exactly is shackling its development?
1. Generative AI landing difficulties: how to break the “wall” between the basic model and developers?
No one wants to miss the generative AI wave. However, the current ultra-high barrier to entry for generative AI has blocked most players from the door. In the past years, model training through “deep learning + big computing power” is the most mainstream technical way to realize artificial intelligence. But the commercialization of the big model landing, must first return to the costing.
First of all, the big model is a huge demand for computing power, which is a huge “gold swallowing beast”. the cost of training GPT-3.5 model is about 3 million to 4.6 million dollars, and the training cost of some bigger language models is even as high as 12 million dollars. Self-developed models are a “bottomless pit” that startups without strong financial strength cannot afford. In addition, generalized models can’t solve all the problems, and can help enterprises accomplish very limited things. The training of large models are based on public data on the Internet to complete, and many products are relatively isolated without forming a coherent, holistic workflow, do not have the ability to customize. It means that developers need to do a lot of personalized debugging with private data, and the threshold of development and training is extremely high. And because of the huge investment in the early stage, even after the commercialization of large models, it often takes a long time to achieve profitability. Therefore, in order for generative AI technology to really land in all walks of life to play a role, there is an urgent need for an affordable, efficient, low-threshold solution to allow more people to participate in the development of generative AI. Then, how can we bridge the gap between the basic model and the end application? At present, the cloud platform, which provides one-stop AI professional hosting services, may be the best solution path at present.
The cloud platform has sufficient and flexible computing power resources, small and medium-sized enterprises do not have to buy and maintain expensive hardware equipment to meet personalized development needs. Users can conveniently invoke third-party resources and outsourcing services on the cloud platform through APIs and SDKs, seamlessly connecting their applications and services to the cloud platform and maximizing the simplification of the development process. In addition, cloud platforms can help solve data privacy and security issues. Over the past few months, many large enterprises, including Apple, Samsung, TSMC, and Bank of America, have introduced relevant policies to explicitly prohibit employees from using ChatGP, and have begun to self-research large models. For small and medium-sized enterprises that don’t have the strength to self-research, choosing a cloud platform that can provide security measures including data encryption, authentication, and compliance tools is a great low-cost option. In response to the current wave of generative AI, do cloud platforms already have considerable capabilities for large model development and are able to provide full-process services for generative AI? At the recently concluded Amazon CloudTech New York Summit, we saw a complete cloud-based solution for generative AI.
2. Amazon Cloud Technology, Creating a New Paradigm for Generative AI Inclusion
This time, Amazon Cloud Technology continues its consistent “pragmatic” style, aiming at the current pain points faced by the transformation of generative AI applications, and uploading a series of new functions and services. From hardware to software, from the development side to the application side, trying to build a generative AI service platform with the most complete functions and the strongest capabilities.Amazon Bedrock service: build a generative AI development “fast track” for the development level of the basic model training cost is expensive, the environment is complex to deploy the problem of April this year, the Amazon Cloud Technology first announced the launch of the Amazon Bedrock service. Amazon Cloud Technology first announced the launch of Amazon Bedrock service, which allows users to use APIs to easily access basic models from different vendors through scalable, reliable and secure Amazon Cloud Technology hosting services, and use them to build generative AI applications.
At the time, in addition to its own Titan grand model, premiere third-party partners and base models included AI21 Labs’ Jurassic-2, Anthropic’s Claude, and Stability AI’s Stable Diffusion.At this New York Summit, Amazon announced that it was once again adding former generative AI field Cohere, one of the largest unicorns in the field, as a vendor, and also added new foundational models including Claude 2, Anthropic’s latest language model, and Stability AI’s latest version of its suite of literate graphical models, Stable Diffusion XL 1.0.
Amazon Bedrock believes that in the future, there will not be one model that governs everything, and by continuously integrating the industry’s most leading base models, users will be able to conveniently invoke the most appropriate model according to their needs. But after the basic model, there is still a thorny problem that has not been solved – how to use these models for personalized application development? The cloud platform has to further address the issues of private data learning, system integration and debugging, and automated task execution.
Let’s take an example of e-commerce returns that we often encounter in our daily lives. You bought a pair of shoes in the e-commerce platform is not very satisfied want to find customer service to change a color, if this time the customer service is ChatGPT and other general chatbots, how will he answer you? — “Sorry, my training data cutoff date is September 2021 and there is no information about this pair of shoes.” In order to make the big model really work, the first thing to do is to “feed” all the information related to the shoes to the model in advance, including the model color of the shoes, the platform’s return policy, inventory information, etc., so that the model can accurately give feedback. While giving the information, the AI also needs to be able to chat while performing all the operations related to the exchange of goods in an orderly and safe manner in the background. In the past this would have been a huge undertaking for developers, but now Amazon has a new service called Amazon Bedrock Agents that puts it all within reach.
The latest Amazon Bedrock Agents service is able to build on the base model by packaging the definition of conversations, the acquisition and parsing of information external to the model, API calls, task execution, etc. into a fully managed service that allows for timely and targeted output. In this way, developers do not have to spend huge amounts of money to develop their own basic models from scratch, nor do they have to spend a lot of time and manpower to carry out personalized deployment and debugging of the model, thus enabling developers to put more energy into the construction and operation of AI applications, so that small and medium-sized developers who do not have strong financial and technological strength can join in the generative AI wave.
“Vector data + hardware computing power” dual escort, casting application development of the strongest brain + the strongest base for the custom development of the model, in addition to the need for professional hosting services such as Amazon Bedrock, but also need to compute, storage, security and other related capabilities to ensure that the model continues to be available, and iterative upgrades. Undoubtedly, data is the substrate for the emergence and development of AI. In order to learn and understand the complexity of human language, generative AI needs a large amount of training data, which usually exists in the form of “vectors”, i.e., natural language is transformed into numbers that can be understood and processed by computers.
So what is vector data and why is it critical to the development of generative AI? Suppose you are using a music recommendation software, we can quantitatively label each song according to three features such as rhythm, lyrics, melody, etc., for example, the first song is (120, 60, 80), the second song is (100, 80, 70), when you tell the system that you like the rhythm of the first song, the system will find the rhythmic vector data of this song “120”, looks up other vectors in the database that are similar to this vector, and then recommends songs with similar characteristics to you. Of course, there’s more than just three dimensions, and a piece of data can be labeled into more latitudes. In natural language processing, “word vectors” represented by word embeddings are typically hundreds of dimensions, while in image processing, image vectors represented by pixel values can have thousands to millions of dimensions. The “vectorized” data will be stored in vector databases to efficiently retrieve and generate the most relevant or similar data in high-dimensional spaces.
However, vectorizing and storing data is not an easy task, and is often labor-intensive and time-consuming. To address this issue, Amazon Cloud Technology this time introduced a vector engine for Amazon OpenSearch Serverless, which is capable of supporting simple API calls for storing and querying billions of Embeddings (the process of mapping high-dimensional data into a low-dimensional space). Amazon Cloud Technology also said that all Amazon Cloud Technology databases will have vector capabilities in the future, becoming the “strongest brain” for developers at the AI data level.
In addition to vector engine support, in the arithmetic level, Amazon Cloud Technology has also been committed to building low-cost, low-latency cloud infrastructure. Amazon Cloud and NVIDIA have been working together for more than 12 years, providing large-scale, low-cost GPU solutions for a variety of applications such as artificial intelligence, machine learning, graphics, gaming, and high-performance computing, and have unrivaled experience in delivering GPU-based instances. This time, Amazon Cloud Technology demonstrated its latest P5 instances powered by NVIDIA’s H100 Tensor Core GPUs, enabling lower latency and efficient horizontal scaling performance.
The P5 instance will be the first GPU instance to utilize Amazon Cloud Technology’s second-generation Amazon Elastic Fabric Adapter (EFA) networking technology. Compared to the previous generation, P5 instances can reduce training time by up to six times, from days to hours, a performance improvement that will help customers reduce training costs by up to 40 percent. With the second generation of Amazon EFA, users are able to scale their P5 instances to more than 20,000 NVIDIA H100 GPUs, providing the supercomputing power needed by customers of all sizes, including startups and large enterprises.
Lowering the Barrier to Generative AI and Maximizing User Empowerment with Products. In addition to tools and platforms for generative AI development, enterprises need ready-to-use generative AI products in their daily operations to help improve work and management efficiency. In this regard, Amazon Cloud Technology has also launched a number of products that can be used directly in work scenarios, and these products cover both the underlying developers and a large number of non-technical staff in the enterprise.
For example, in the field of code development, since Amazon Cloud Technology in June last year for the first time after the launch of the AI programming assistant Amazon CodeWhisperer, the function has now become one of the daily essential tools for many developers. Amazon CodeWhisperer is based on billions of lines of open source code training, and can be based on the code comments and the existing code in real time to generate code suggestions, in addition to security vulnerability scanning. It currently supports 15 programming languages including Python, Java, and JavaScript, and integrated development environments including VS Code, IntelliJ IDEA, JupyterLab, and Amazon SageMaker Studio.
To further improve development productivity, at the New York Summit, Amazon Cloud Technologies officially announced that Amazon CodeWhisperer is also supported by Amazon Glue Studio Notebooks, which allows developers to write specific tasks in natural language, followed by the Amazon CodeWhisperer. With Amazon Glue Studio Notebooks, developers can write a specific task in natural language, and then Amazon CodeWhisperer can suggest one or more code snippets that accomplish the task directly in Notebooks for developers to use and edit. And for non-developer work scenarios, by combining Amazon Bedrock’s big language modeling capabilities with Amazon QuickSight Q, which supports natural language Q&A, users are provided with a new service for business intelligence based on generative AI.
For example, if you are a financial analyst, you can use natural language to give commands like chatting with ChatGPT, and in a few seconds Amazon QuickSight Q can complete the operation of searching for key financial information or creating a visual chart of the company’s finances, and at the same time, it can help you summarize the characteristics of the trend and make recommendations. Similar ready-to-use products include Amazon Entity Resolution, which helps organizations break down internal information silos and accelerate data-driven decision-making, and Amazon HealthScribe, which helps healthcare software vendors easily build clinical applications based on generative AI, expanding the use of generative AI products in a variety of industries. The following are some examples of the ways in which generative AI products are expanding their use in various industries
3. Unleash the “Cloud Power” of the AI Era
The development of generative AI requires the cloud, and more importantly, a large number of cloud-based tools and services. After the big model, the next stage of generative AI technology will definitely develop in the direction of diversity and personalization, and we can see both more general productivity tools, but also see a variety of AI products targeting specific scenarios. And in this process, the cloud platform will play an increasingly critical role.
On the one hand, the cloud platform will greatly reduce the threshold of AI application development. With the support of the platform’s arithmetic power and basic models, developers basically don’t need to care about hardware and infrastructure issues, so they can put more time and energy into business and operations. On the other hand, the cloud platform can continuously accelerate the development and operational efficiency of AI applications. Users can develop and manage applications by directly calling APIs, and share them among teams or organizations safely and conveniently.
With the help of the cloud platform, the future of generative AI will no longer be just a “money-burning game” that only giants can play, but more ordinary people will be able to sit at the poker table. As one of the industry leaders in the field of cloud services, Amazon Cloud provides more than 200 services covering a wide range of areas such as computing, storage, database, network, developer tools, security, analytics, IoT, enterprise applications, etc., and its cloud infrastructure covers the world. At the same time, Amazon Cloud Technology is also a leader in the field of artificial intelligence and machine learning, and has been continuously providing and updating a series of end-to-end AI-related services for years, allowing developers to develop and deploy generative AI applications flexibly, conveniently, and at low cost.
This time, Amazon Cloud Technology released generative AI tools “family bucket”, the core purpose is to further reduce the threshold of generative AI development, so that more ordinary people who do not understand the big model, do not understand artificial intelligence can also quickly join the development and application of generative AI. The importance of generative AI does not lie in how big the model is and how strong the ability is, but more importantly, how it can evolve from the basic model into specific applications in various fields, thus empowering the development of the entire economy and society. Now, Amazon Cloud Tech is becoming that bridge.