What is Generative AI?

Like all artificial intelligence, generative AI works by using machine learning models—very large models that are pretrained on vast amounts of data.

Foundation models

Foundation models (FMs) are ML models trained on a broad spectrum of generalized and unlabeled data. They're capable of performing a wide variety of general tasks.

FMs are the result of the latest advancements in a technology that has been evolving for decades. In general, an FM uses learned patterns and relationships to predict the next item in a sequence.

For example, with image generation, the model analyzes the image and creates a sharper, more clearly defined version of the image. Similarly, with text, the model predicts the next word in a string of text based on the previous words and their context. It then selects the next word using probability distribution techniques.

Large language models

Large language models (LLMs) are one class of FMs. For example, OpenAI's generative pre-trained transformer (GPT) models are LLMs. LLMs are specifically focused on language-based tasks such as such as summarization, text generation, classification, open-ended conversation, and information extraction.

Read about GPT »

What makes LLMs special is their ability to perform multiple tasks. They can do this because they contain many parameters that make them capable of learning advanced concepts.

An LLM like GPT-3 can consider billions of parameters and has the ability to generate content from very little input. Through their pretraining exposure to internet-scale data in all its various forms and myriad patterns, LLMs learn to apply their knowledge in a wide range of contexts.

Traditional machine learning models were discriminative or focused on classifying data points. They attempted to determine the relationship between known and unknown factors. For example, they look at images—known data like pixel arrangement, line, color, and shape—and map them to words—the unknown factor. Mathematically, the models worked by identifying equations that could numerically map unknown and known factors as x and y variables.

Generative models take this one step further. Instead of predicting a label given some features, they try to predict features given a certain label. Mathematically, generative modeling calculates the probability of x and y occurring together. It learns the distribution of different data features and their relationships.

For example, generative models analyze animal images to record variables like different ear shapes, eye shapes, tail features, and skin patterns. They learn features and their relations to understand what different animals look like in general. They can then recreate new animal images that were not in the training set.

Next, we give some broad categories of generative AI models.

Diffusion models

Diffusion models create new data by iteratively making controlled random changes to an initial data sample. They start with the original data and add subtle changes (noise), progressively making it less similar to the original. This noise is carefully controlled to ensure the generated data remains coherent and realistic.

After adding noise over several iterations, the diffusion model reverses the process. Reverse denoising gradually removes the noise to produce a new data sample that resembles the original.

Generative adversarial networks

The generative adversarial network (GAN) is another generative AI model that builds upon the diffusion model’s concept.

GANs work by training two neural networks in a competitive manner. The first network, known as the generator, generates fake data samples by adding random noise. The second network, called the discriminator, tries to distinguish between real data and the fake data produced by the generator. 

During training, the generator continually improves its ability to create realistic data while the discriminator becomes better at telling real from fake. This adversarial process continues until the generator produces data that is so convincing that the discriminator can't differentiate it from real data.

GANs are widely used in generating realistic images, style transfer, and data augmentation tasks.

Variational autoencoders

Variational autoencoders (VAEs) learn a compact representation of data called latent space. The latent space is a mathematical representation of the data. You can think of it as a unique code representing the data based on all its attributes. For example, if studying faces, the latent space contains numbers representing eye shape, nose shape, cheekbones, and ears.

VAEs use two neural networks—the encoder and the decoder. The encoder neural network maps the input data to a mean and variance for each dimension of the latent space. It generates a random sample from a Gaussian (normal) distribution. This sample is a point in the latent space and represents a compressed, simplified version of the input data.

The decoder neural network takes this sampled point from the latent space and reconstructs it back into data that resembles the original input. Mathematical functions are used to measure how well the reconstructed data matches the original data.

Transformer-based models

The transformer-based generative AI model builds upon the encoder and decoder concepts of VAEs. Transformer-based models add more layers to the encoder to improve performance on text-based tasks like comprehension, translation, and creative writing.

Transformer-based models use a self-attention mechanism. They weigh the importance of different parts of an input sequence when processing each element in the sequence.

Another key feature is that these AI models implement contextual embeddings. The encoding of a sequence element depends not only on the element itself but also on its context within the sequence.

How transformer-based models work

To understand how a transformer-based models work, imagine a sentence as a sequence of words.

Self-attention helps the model focus on the relevant words as it processes each word. To capture different types of relationships between words, the transformer-based generative model employs multiple encoder layers called attention heads. Each head learns to attend to different parts of the input sequence. This allows the model to simultaneously consider various aspects of the data.

Each layer also refines the contextual embeddings. The layers make the embeddings more informative and capture everything from grammar syntax to complex semantic meanings.

If your organization wants to implement generative AI solutions, consider the following best practices to enhance your efforts.

Begin with internal applications

It’s best to start generative AI adoption with internal application development, focusing on process optimization and employee productivity. You get a more controlled environment to test outcomes while building skills and understanding of the technology. You can test the models extensively and even customize them on internal knowledge sources.

This way, your customers have a much better experience when you do eventually use the models for external applications.

Enhance transparency

Clearly communicate about all generative AI applications and outputs, so your users know they are interacting with AI and not humans. For instance, the AI can introduce itself as AI, or AI-based search results can be marked and highlighted.

That way, your users can use their own discretion when they engage with the content. They may also be more proactive in dealing with any inaccuracies or hidden biases the underlying models may have because of their training data limitations.

Implement security

Implement guardrails so your generative AI applications don't allow inadvertent unauthorized access to sensitive data. Involve security teams from the start so all aspects can be considered from the beginning. For example, you may have to mask data and remove personally identifiable information (PII) before you train any models on internal data.

Test extensively

Develop automated and manual testing processes to validate results, and test all types of scenarios the generative AI system may experience. Have different groups of beta testers who try out the applications in different ways and document results. The model will also improve continuously through testing, and you get more control over expected outcomes and responses.

Amazon Web Services (AWS) makes it easy to build and scale generative AI applications for your data, use cases, and customers. With generative AI on AWS, you get enterprise-grade security and privacy, access to industry-leading FMs, generative AI powered applications, and a data first approach.

Choose from a range of generative AI technologies that support all types of organizations in every stage of generative AI adoption and maturity:

• Code generation is one of the most promising applications for generative AI. With Amazon CodeWhisperer, an AI coding companion, you can get great results in developer productivity. During the preview, Amazon ran a productivity challenge. Participants who used CodeWhisperer were 27 percent more likely to complete tasks successfully. On average, they completed tasks 57 percent faster than those who didn't use CodeWhisperer.
• Amazon Bedrock is another fully managed service that offers a choice of high-performing FMs and a broad set of capabilities. You can easily experiment with various top FMs, privately customize them with your data, and create managed agents that execute complex business tasks.
• You can also use Amazon SageMaker JumpStart to discover, explore, and deploy open source FMs—or even create your own. SageMaker JumpStart provides managed infrastructure and tools to accelerate scalable, reliable, and secure model building, training, and deployment.
• AWS HealthScribe is a HIPAA-eligible service empowering healthcare software vendors to build clinical applications that automatically generate clinical notes by analyzing patient-clinician conversations. AWS HealthScribe combines speech recognition and generative artificial intelligence (AI) to reduce the burden of clinical documentation by transcribing patient-clinician conversations and generating easier-to-review clinical notes.
• Amazon Q in QuickSight helps business analysts easily create and customize visuals using natural-language commands. The new Generative BI authoring capabilities extend the natural-language querying of QuickSight Q beyond answering well-structured questions (such as “what are the top 10 products sold in California?”) to help analysts quickly create customizable visuals from question fragments (such as “top 10 products”), clarify the intent of a query by asking follow-up questions, refine visualizations, and complete complex calculations.

Get started with generative AI on AWS by creating an account today.