Firstly some clarity AI is what happens, LLM (large language model) is how. AI has rapidly become more prevalent over the past few months despite not being a new concept at all (the Ferranti Mark 1 was run in 1951). Let’s go through the types of AI.
Deterministic AI
These are systems that don’t have memory and respond to tasks eg providing purchase history info.
Generative AI
Generative AI The most common use case right now. These systems have some element of memory or more specifically are trained on large data sets. ChatGPT is one of the most famous examples of this. GPT stands for Generative Pre-Trained Transformer, the learning mechanism behind it uses supervised learning and is supposed to enhance this learning based on human feedback.
Human Mind
This is an approach to understand human feelings and emotions. An example of this is a chatbot being used for interaction about emotional issues with the purpose of reducing depression.
Self Aware
Skynet in the Terminator movies is a very famous example of this. We are a long way away from this as a reality.
AWS services
Amazon Textract
- Amazon Textract is an approach to solving Optical Character Recognition problems. Traditionally OCR was never great at recognising things like forms in documents, Textract uses an LLM to recognise and interpret these values based on context.
- PDF, Word and image documents are all supported.
- Example use cases include healthcare patient forms, legal documents such as mortgage applications and financial invoice documents.
- Textract works in a pipeline fashion so it can scale up or down as demand requires.
Read about a real Australian healthcare use case
Transcribe
- Transcribe is tool for the purpose of converting audio to text, using LLM for training on specific voices and custom vocabularies.
- English, Spanish, French, German, Portuguese and Mandarin are the main supported languages.
- Transcripts can redact personal data such as credit card information so compliance is no issue.
- Further insights into transcripts can be formed as integration with Amazon Comprehend (a Natural Language Processing tool) is natively supported.
- Real time transcription is supported as transcriptions are uploaded to S3 and then processed via web sockets.
Sagemaker
Now we get to services that are not really AI services. Why have you included them then I hear you cry! Remember earlier where I explained that AI is what happens and LLM is how it happens?…. Well, here’s a great way to train an LLM, admittedly use cases for this are slightly rarer due to most people wanting to just use something that already does a job for them. Sagemaker Ground Truth is integrated for detailed data labelling to annotate images, text classification and audio transcriptions.
Here’s a Youtube video demonstrating an example using Sagemaker to analyse healthcare images
Bedrock
Bedrock allows you to use existing LLM’s and specify which is most appropriate for your use case. Each LLM is built for a slightly specific use case, such as text generation or multilingual support. Different models can be evaluated and switched between seamlessly.
Read about how United Airlines used Bedrock to analyse customer data
Amazon Q
- This is not Amazon’s take on James Bond characters, unfortunately. Q is a generative AI service for the purpose of natural language querying and question answering.
- It can be used in conjunction with S3, databases and web crawlers.
- Use cases often include using Amazon Q to answer questions based on product documentation.
- Business Intelligence can be assisted with Q’s integration to Amazon Quicksight, a tool for creating dashboards.
Read about how Q was used to assess product runbooks and show dependencies
