Generative artificial intelligence under AI innovation—Amazon Bedrock

Table of contents

When is generative artificial intelligence

Generative Artificial Intelligence (Generative AI for short) is a category based on machine learning and artificial intelligence technology. Its goal is to enable computer systems to autonomously generate various types of data, such as text, images, audio, etc. , rather than just imitating or classifying existing data. The core capability of generative AI is to create new content rather than just re-applying known patterns. Among them, generative artificial intelligence models are an important part of generative AI, and chatGPT is one of the representative models. The principle of Chengdu AI is based on deep learning, especially neural network technology, which can learn the distribution and patterns of data by analyzing a large amount of training data, and then use these learned patterns to generate new data. Generative AI can not only generate high-quality text, but can also be used in various fields such as image generation, music creation, and video generation.

Amazon Cloud’s Generative AI Products

AWS offers many possibilities for developers of large-scale language models. Amazon Bedrock is the easiest way to build and scale generative AI applications using LLM. Amazon Bedrock is a fully managed service that provides LLMs from Amazon and leading AI startups via API, so you can choose from a variety of LLMs to find the best model for your use case.

Amazon Bedrock VS Amazon SageMake

mazon Bedrock and Amazon Sagemaker are two separate services
Amazon SageMaker is an end-to-end machine learning platform. Its functions include end-to-end capabilities from data annotation to data training to deployment, continuous monitoring after launch, and iteration based on original data. Before the emergence of large generative AI models, it was already very mature in helping customers solve problems.
Amazon Bedrock is positioned to solve the following core problems:
First, it can help customers quickly find industry-leading models and access them through APIs as much as possible without customers having to manage the underlying hardware and operation and maintenance themselves.
Second, when users use Amazon Bedrock for model tuning and training, the difference from using Amazon SageMaker is that Amazon SageMaker is first for data scientists, who need to write code, provide data, and add parameters for model and tuning. In terms of application form, in Bedrock, customers only need to provide 20 labeled data, which can be quickly deployed without writing too much code.
The basic difference between the two is to draw user portraits. SageMaker is aimed at professionals, while Bedrock is designed to lower the threshold as much as possible so that users in industry-focused scenarios can use it better.
As a new, generative AI production tool, Bedrock relies on the existing five preset basic models and directly calls through API to generate more content. Amazon SageMaker is a comprehensive, full-featured machine learning tool, including data annotation to training to inference. It has all the capabilities, and it also has some preset capabilities. You can choose applicable scenarios as needed. In other words, Bedrock is only in the field of generative AI, and Amazon SageMaker is for all machine learning and artificial intelligence fields. Generative AI is only a part of AI, while Amazon SageMaker is for the entire AI/ML.
In addition, in terms of deployment integration, Amazon SageMaker and Bedrock have some combinations. In fact, some functions of SageMaker can be reused on models trained by Bedrock. For example, if a customer customizes a model through Bedrock, when the new model is applied, the model management process can be interoperable. Customers can integrate basic models with Amazon SageMaker machine learning capabilities, use Experiments to test different models, use Pipelines to manage basic models at scale, and more.

The origin of Amazon Bedrock

Specifically, Bedrock mainly consists of two parts, one is Amazon Cloud Technology's own model Titan, and the other is the basic model from startups AI21 Labs, Anthropic, and Stability AI.

The basic model specifically includes:

Amazon Jiken Titan
Claude（Anthropic）
Jurassic-2（AI21 Labs ）
Stable Diffusion（Stability.AI）

The construction of Titan's basic model is based on Amazon Cloud Technology's more than 20 years of experience in the field of machine learning. Titan contains two large language models, one is Titan text for generating text, and the other is Titan Embeddings for personalizing web searches. Titan text targets tasks such as summarization, text generation, classification, open question and answer, and information extraction. The text embedding Titan Embeddings model can translate text input (words, phrases, large articles) into digital expressions containing semantics (embeddings into encoding).

Users can customize Titan models with their own data. Moreover, Amazon Cloud Technology protects user data privacy very much and will not use user data to retrain the Titan model. Moreover, unlike the "illusions" that often occur with other large models, Titan pays great attention to accuracy during training to ensure that the responses generated must be of high quality. In addition to Amazon Cloud Technology’s Titan model, developers can also take advantage of other basic models. These include the Jurassic-2 multilingual large language model series developed by AI21 Labs, which can generate text content based on natural language instructions and currently supports Spanish, French, German, Portuguese, Italian and Dutch. There is also the large language model Claude developed by Anthropic, which can perform multiple rounds of dialogue and text processing tasks. The third basic model is Stability AI’s text image generation model Stable Diffusion. Through these models, developers can customize their own models with one click using only 20 samples.

For example, if a marketing manager wants to develop advertising creative for a new handbag product, he only needs to provide Bedrock with the best annotated ads and new product descriptions, and Bedrock can automatically generate media tweets, display ads, and product web pages. Again, all data is encrypted and no customer data is used to train the underlying model. Currently, partners such as Coda AI, Deloitte, Accenture, and Infosys have used Bedrock.

Application and use of Amazon Bedrock

First, before using Bedrock, you must first apply for model access permissions. Currently, it is only available in us-east-1, us-west-2, ap-southeast-1, ap-northeast-1. You need to apply for model access permissions first. For all models, they are not open by default. In the model access interface, first select Request Access, which will ask you to fill in the company name, website, and purpose, and then request access permission. Try to choose overseas company URLs here. Currently, China is unstable. In addition, the payment company for this account also needs to be selected overseas, and then choose the model you need.
After logging in to the Bedrock console, first go to the model access interface, click Edit, and check the required models. If you do not need the model, just leave it blank and it will be automatically deleted. Then select Save. After saving, you can use the provided model. Such as Claude and Stable Diffusion XL, etc. Because when using overseas, the console defaults to English.

After creation, you can see many functions through the left side of the console. We can select our base model. In the Amazon Bedrock console, we can group it by various model attributes. You can also filter model views, search for models, and view information about model providers.

After selecting the model, open it in the playground and you can use the various functions of the model.

Amazon Bedrock use cases

For example, if we make our own AI cognitive chat tool, we can choose some of Bedrock's basic models or our own fine-tuned models, such as: