Amazon Bedrock

AWS unveiled the Amazon Bedrock that can offer a way to create applications that are powered by Generative AI, through AI21 Labs, Anthropic & Stability AI. This is available in a “limited preview” and lets users access Titan FMs, a family of models trained in-house by AWS.

It offers access mainly to the ultra-large ML models on which Generative AI depends. This opens up an array of FMs from leading providers. Therefore, AWS customers are capable of using the best models for particular requirements. If you are willing to build and scale Generative AI apps with foundation models, you should choose Amazon Bedrock.

What do you know about Amazon Bedrock?

Amazon Bedrock is actually a fully-managed service that creates FMs from Amazon and leading AI startups available via an API. You can select from different foundation models to find the best suitable one for your use case.

Users can start their work, experiment with foundation models, customize these with their data and deploy these into apps with the help of AWS tools. This service’s agents are fully managed. They help developers to produce generative-AI apps that can provide up-to-date answers depending on proprietary knowledge sources.

How does it work?

Its serverless experience allows you to find the correct model to start the process and customize FMs with your data. Then, it integrates & deploys these into apps with the help of the AWS tools and the capabilities with which you are familiar with ( including integration with Amazon SageMaker ML features, such as Experiments for testing various models & Pipelines in order to manage foundation models at scale) without managing any infrastructure.

What makes Amazon Bedrock different?

Amazon’s largest cloud computing competitors provide similar services for creating generative AI apps. So, now the question will be why a company will choose to work with Amazon. It is necessary to know that Amazon is not currently developing consumer-facing generative AI apps. Microsoft has invested in OpenAI, which is the company behind ChatGPT & Dall-E. Alphabet (GOOG -1.07%) (GOOGL -1.19%) is developing its AI-powered chatbot Bard using its LaMDA model.

Amazon lets users access Titan, its AI model. Developers who are interested in using Amazon’s models can use AWS. The company has developed its chips to train AI models & produce inferences like responses to prompts, dubbed Trainium & Inferentia, respectively.

The design of the chips allows you to decrease the pressure of workloads & save money on calculation. Therefore, developers can save more money if they use AWS. You should know that Google and Microsoft have their chip designs for AI training.

What can Amazon Bedrock do?

This ML platform is similar to Amazon SageMaker. Although both services include many sub-components, the latter is more complicated. Besides, ML engineers use this to build & train custom ML models. Amazon Bedrock has a more user-friendly platform that can be used to build & scale generative AI apps via the use of FMs.

Amazon Bedrock is used in different use cases:

Text generation:

It enables you to create new and original written content such as short stories, social media posts, articles, web page copy, or school essays.

Chatbots:

This service can generate conversational interfaces, including chatbots & virtual assistants, for clients with the intention of improving user experience. The platform can integrate directly with Amazon Lex, which is a chatbot service in AWS.

Search:

Bedrock enables you to search, find or synthesize information so that you are able to answer questions from the library of data.

Text summarization:

It lets you summarize textual content like blog posts, essays, books, documents, etc. Therefore, users don’t have to read the full content in a verbose way.

Image generation:

This service produces an artistic image of various subjects, environments, & scenes.

Personalization:

It allows you to personalize the way of dealing with your customer. This one provides more relevant & contextual product recommendations. It is far better than keyword matching.

How can you access and try the new Amazon Bedrock service?

This service is not available yet in the AWS Management Console. However, users can show their interest by filling out the form: https://pages.awscloud.com/generative-AI-interest-learn.html.

Users must have an AWS Account ID to sign up for service updates.

Everyday Use Cases for Amazon Bedrock:

It can be used to develop multiple AI applications like predictive maintenance, natural language processing, etc. These are some of the use cases:

Predictive Maintenance:

Using this service, it is possible to develop machine learning models for predicting when the equipment is likely to fail. Thus, it enables businesses to schedule maintenance before any breakdown.

Fraud Detection:

It can develop models to detect fraud in financial transactions. Thus, it can help companies to detect fraudulent activity.

Natural Language Processing:

The service produces models to analyze & interpret natural language. Thus, businesses can automate customer service & support.

Image Recognition:

The service allows users to produce models used for analyzing & interpreting images. Thus, it enables companies to automate tasks like quality control in manufacturing.

Integrations:

The service integrates with several software tools:

AWS or Amazon Web Services for database storage, compute power, content delivery, etc.
Anthropic’s Claude AI for building & processing human-like text
Stability AI uses augmented technology & collective intelligence to implement solutions.
Stable Diffusion for generating realistic pictures.
Amazon Titan for accessing foundation models through an API

Features of Amazon Bedrock:

A Wide Variety of FMs:

Amazon Bedrock customers can get a huge variety of accessible & advanced foundation models:

Claude: Anthropic’s LLM is used to perform many conversational tasks and many test processing.

Jurassic-2: It is from AI21 Labs, which uses natural language commands to make special text in German, French, Spanish, Italian, Dutch, & Portuguese.

Stable Diffusion: Stability AI like this enables you to access many text-to-image FMs. In this way, you can get realistic and uniquely designed images.

Amazon Titan: It lets you access several FMs for creating text and photos. This one has two new LLMs.

Choose any of these and get started with a project, whether it is image and text creation or app development.

Titan FMs:

The company has been previewing the latest Titan FMs before these become available widely. Initially, they have two Titan FMs:

Generative LLM: It is used for text generation, text summarization, open-ended Q&As, information extraction, and classification.

Embeddings LLM: This one is able to translate text inputs into numerical representations. For instance, it can translate large text units, phrases, words, etc., into embeddings or numerical representations which have the semantic meaning of the text.

Although LLM is not used to generate text, people use this for search, personalization, etc. Comparing embeddings can enable the models to create more contextual & relevant responses and helps to find products easily & quickly.

Customization:

A high customization level is offered by the service. Using it, you can easily customize an AI model with your data to make this suitable for your project. Hence, users have to point out some labelled examples in S3. Thus, it helps to fine-tune your model for the particular use case. You can do the job with 20 labelled examples. While it helps to eliminate the need to annotate large data volumes, it can save effort & time.

Example: Suppose your profession is content marketing at a cloth brand. So, you need to create a campaign copy to attract potential customers toward an upcoming line of shirts.

This service is beneficial in this case. Amazon Bedrock can be provided with a few examples of the best-performing campaign copies & descriptions from the past. After that, the service will generate another private copy of FM which buyers can access only. Then, they train the model to let this automatically produce useful campaign copy for new shirts.

Security and Privacy:

In order to train base models, the service does not use customer data. Besides, it can encrypt all data and does not leave a customer’s VPC or a Virtual Private Cloud. Thus, the service can strive to maintain customer trust. Customers can be assured that the data remains protected.

Additionally, the design of Amazon’s Titan FMs enables you to identify harmful data quickly and remove this. It is able to find out wrong content in a user’s input and reject this. Besides, it filters the AI model’s output which contains inappropriate content such as violence, profanity, hate speech, etc.

Accessibility:

The service offers the benefits of greater accessibility of FMs to every small, mid-sized, and big business or enterprise. You might experience the power of foundation models throughout your organization. Additionally, it is possible to accelerate machine learning usage & empower developers to create your generative AI apps. Infosys, Accenture, Deloitte and many more companies are developing practices so that companies can move faster in generative AI usage.

Scalability:

AWS can offer you a scalable & more trustworthy experience that develops modern AI apps. It is possible to integrate the selected & customized foundation models into scalable apps. And, you can deploy these quickly using the tools given by AWS and that you use.

As a result, you do not need to manage any infrastructure. For instance, there is no need to manage integrations with SageMaker ML functionalities for experimenting with different models, Pipelines to handle foundation models at scale. If data is stored on AWS already, it can easily scale your data & use generative AI with Amazon Bedrock with better security.

Use Case Implementation & Projected Costs:

We have given an example of a sample cost breakdown, which depends on a fraud detection use case, an HR platform, and absence management that supports 100,000 lives. Besides, it supports 15,000 claims yearly and 10,000 driver’s licenses on average to verify identity annually and 10,000 health provider paperwork which can be scanned for physician approval per year so that it can be identified how many claims have submitted invalid licenses and how many provider paperwork signals fraud.

Implementation Steps:

Data Collection: It is required to collect absence management data, driver’s license data, & health provider paperwork data. This absence management data contains information on 100,000 employees & their claims. The driver’s license data as well as the health provider paperwork data need to be scanned annually.

Data Cleaning: For confirming high-quality data for model training, it is required to clean the data & remove duplicates or irrelevant information.

Data Labeling: Hence, it is necessary to label data for model training depending on detecting invalid licenses & health provider paperwork signaling fraud.

Model Training: It is possible to train a fraud detection model with Amazon Bedrock’s default components for image classification & natural language processing.

The Model Testing & Validation: A small data set can be used to test the model and validate its accuracy before the deployment in a production environment.

Model Deployment: It is possible to deploy FM to a cloud environment & produce an API endpoint with the intention of receiving claiming data & identifying fraud.

Continuous Monitoring & Improvement: It is essential to monitor the model’s performance and improve it depending on feedback & new data.

Projected Costs:

Data Storage: Around $34 monthly is the expected price of data storage for absence management data, driver’s license data, and health provider paperwork data. Hence, the cost in detail is as follows:

$4 monthly for absence management data,
$10 monthly for driver’s license data, and
$20 monthly for health provider paperwork data.

Machine Learning Training:

The fraud detection model’s training cost relies on the model’s complexity and the necessary computing resources. Training costs can be from some hundred dollars to some thousand dollars, as expected for a medium-sized model.

Model Deployment: For an average of 15,000 claims each year, this cost is around $15 per month.

Continuous Monitoring and Improvement: The cost of monitoring and improving the model would depend on the frequency of updates and the amount of data involved. Assuming monthly updates & a moderate amount of data, the cost would be around $50 every month.

Above all, it can be said that the estimated cost to develop a fraud detection model using Amazon Bedrock can be around $400 — $2,500 each month, based on the model’s complexity & the involved data.

The bottom line:

Amazon Bedrock, a powerful AI platform, can offer businesses different services & tools so that they can build, train, & deploy ML models. Besides, businesses can develop AI apps using its scalable infrastructure, pre-built algorithms, & data management system. It develops multiple apps,predictive maintenance, natural language processing, and so on.