"Large models are only one part of customer needs, but far from all. Customers also need a wider range of basic capabilities. Amazon Cloud Technology has launched self-developed chips, generative AI service Bedrock, and large model Titan, all of which are committed to promoting AIGC. The inclusiveness of technology can lower AIGC’s technical and capital thresholds, making it easier and more cost-effective for more people to use AIGC.”
Recently, Chen Xiaojian, general manager of the product department of Amazon Cloud Technology Greater China, said this at a media communication meeting.
On April 13, Amazon Cloud Technology officially entered AIGC, launched AIGC service Bedrock and its own basic model Titan, and AI programming assistant Amazon CodeWhisperer, and announced that the latest instance based on self-developed training and reasoning AI chips is officially available.
In the field of AIGC, Amazon Cloud Technology definitely has the strength to make basic large-scale models and large-scale language models, but it is more focused on how to build a powerful AIGC cloud base and promote the inclusiveness of AIGC technology.
Four recent technological innovations of Amazon Cloud Technology in the field of AIGC
On the cloud, fully managed and easily customized AIGC services
Facing the explosive growth of AIGC technology, what are the core needs of enterprise customers? Earlier, Swami Sivasubramanian, Global Vice President of Amazon Cloud Technology Database, Data Analysis and Machine Learning, stated in a signed article:
“Customers have told us their main needs now: first, to directly find and access high-performance base models; second, to seamlessly integrate with applications without having to manage large clusters of infrastructure and without adding prohibitive costs .Third, it is easy to get started, based on the basic model, using your own data (more or less) to build differentiated applications. ”
The layout of Amazon Cloud Technology in the AIGC field is based on these core needs.
AIGC service Bedrock is a fully managed service. Users can access the current mainstream AIGC large models through API, including basic models from AI21 Labs, Anthropic, Stability AI, and Amazon's own basic model Titan.
Base models currently supported by Amazon Bedrock
" Fully managed and easy customization are Bedrock's unique strengths. Users don't need to worry about infrastructure details such as instance types, network topology, and endpoints. At the same time, users only need to provide a small number (as low as 20) of labeled instances in Amazon S3. Customize Bedrock's base model for its specific use case." Chen Xiaojian said.
The base model Titan includes two new large language models. Among them, Titan Text focuses on generative NLP tasks, such as writing summaries, creating blogs, text classification, dialogue and information extraction, etc.; Titan Embeddings is used for search and personalization, etc., which can translate text input into embedded codes that contain semantics. Search results are more relevant and contextual, and a similar text embedding model is already used in Amazon.com's product search.
The AI programming assistant CodeWhisperer is now open to all individual developers free of charge. In addition to Python, Java, JavaScript, TypeScript and C#, it supports 10 new development languages such as Go, Kotlin, Rust, PHP and SQL. CodeWhisperer can significantly improve developer productivity, increasing task completion speed by 57% and task success rate by 27% during the preview period.
"We believe generative AI will be a game-changer for developers, so we want it to be available to as many people as possible," Swami said.
"Developers can interact with CodeWhisperer through comments. Generally, when writing code, we will add comments to our own code. In the comments, write "Please help me generate a code for uploading pictures to the cloud", and it will give this code If we accept all the suggestions, we only need to press the Tab key, which is equivalent to accepting the suggestions. Such an interactive experience makes CodeWhisperer more like a smart assistant for developers.” Amazon Cloud Technology Greater China Data Technology Expert Team Director Wang Xiaoye said.
In the era of AIGC, maximize the advantages of the cloud base
The bottleneck of AIGC's development is that computing power has become the consensus of the industry, and more cost-effective computing power is the basis for the rapid development of AIGC.
This time, Amazon Cloud Technology has further upgraded the cost performance of computing power. Currently, the new Trn1n based on the Amazon Trainium chip and the Amazon EC2 Inf2 instance based on the Amazon Inferentia2 chip are generally available.
Trn1 compute instances powered by Trainium can save up to 50% on training costs compared to any other EC2 instance. Taking a two-week training of a large model with trillions of parameters as an example, the GPU server P3dn needs 600 instances, the latest generation GPU instance P4d needs 128 instances, but Trn1 only needs 96 instances.
The Inf2 instance, powered by Amazon Inferentia2, is the only instance built specifically for distributed inference of large Transformer models. Compared with previous-generation instances, Inf2 instances provide up to 4 times higher throughput and up to 10 times lower latency. Compared with GPU-based instances, the performance per watt is improved by up to 45%. It also supports large and complex models such as GPT, and can use a single instance to implement inference of 175 billion parameter models.
At present, many leading AIGC companies are innovating based on the infrastructure of Amazon cloud technology.
In the field of AI painting, the AIGC unicorn Stability AI , which launched the open source AI model Stable Diffusion , is using a large-scale GPU cluster on Amazon Cloud Technology and a high-performance computing cluster composed of Amazon Trainium machine learning training chips to train its generative AI. The basic model, and optimize the cost through the elasticity of model training on the cloud, and finally reduce the training time and cost of open source language models such as GPT-NeoX used by it by 58%.
"Today, the time and money spent on the base model is mostly used for training, because many customers are just starting to deploy the base model into production. However, in the future when the base model enters large-scale deployment, most of the cost will be spent running Model and reasoning. ” When it comes to the demand for computing resources for future training and reasoning, Swami said so.
write at the end
Talking about the unique advantages of Amazon cloud technology in the field of AIGC, Chen Xiaojian said that there are three main points:
First, reduce the difficulty for customers in large-scale training and deployment, and lower the threshold for AI innovation;
Second, provide a variety of model options, including third-party models and self-developed models;
Third, strong security capabilities ensure the data security of customers when training and deploying customized models.
Amazon Cloud Technology believes that the current generative AI model mainly focuses on text and image generation, and is gradually penetrating into audio and video content generation. In the future, there will be more and more cross-modal/multi-modal content generation.
“We are at an exciting inflection point in the mass adoption of machine learning, and we also believe that generative AI will reshape a large number of customer experiences and applications,” concluded Swami.
The pictures in the text are from Photography Network
END
This article is the original work of "Intelligent Evolution".