What do we mean when we talk about GenAI?

A female professional holding an tablet screen

As we explore the potential of generative AI (GenAI), it is important to understand the fundamentals of this exciting technology and how it can create value across a wide range of use cases.

What can GenAI do? While many people most commonly think of GenAI as generating text, it can make much more than that:

  • TextGenAI can generate text in many formats, from poems to emails. It can also translate text between different languages and create answers to user questions.
  • Code: GenAI can write code or find bugs in code. It can help with writing better and optimized code, too. This boosts a coder’s productivity.
  • Audio: GenAI can develop audio files for music, sound clips, podcasts, and audiobooks.
  • Images: GenAI can create images based on the text that users provide describing imagery.
  • Video: Building on how it generates images, GenAI can create a video with the additional complexity of continuity from frame to frame.
  • Synthetic data: GenAI can make synthetic data similar enough to real-world data that it can be used for training machine learning models and testing software.

GenAI helps pair concepts

Recent breakthroughs in GenAI rely on one major concept: the ability to generate new content (such as text, code, or images) similar to the data it was trained on. Generative language models learn about patterns in language and images through training data. Then, given some language or images, the GenAI predicts what comes next.

By learning from the massive amount of information available on the internet, mathematical models can be built to translate one type of paired content to another type of paired content. For example:

  • Questions are often paired with answers.
  • Images are often paired with captions or other descriptions.

Other types of content also commonly occur in sequence, such that GenAI can learn to predict concepts of what should follow certain wording. For example:

  • The word “blog” is commonly followed by the word “post.”
  • A sales meeting agenda typically follows the “introductions” section with a “company overview” section.
  • A list of a software product’s features often lists individual user features, followed by features focused on collaboration, then features focused on security.

Because different types of content commonly have relationships that can be learned by a GenAI model, we can construct a model that returns one type of content based on its predictable relationship to another type of content, such as returning an answer based on a question.

That is what is at play at the root of GenAI.

Foundation models offer building blocks

So, how do we start using these tools? Large language model providers have trained what are referred to as “foundation models” from which we can build.

A foundation model is a machine learning model trained on a vast quantity of data so it can learn patterns of human language and behavior. This allows the models to be easily adapted to perform a wide range of tasks without the need for extensive retraining. We can use prompts to get the desired output.

Prompts include instructions to the large language model on how and what to generate. This might contain the desired format or examples of response styles we find to be ideal. It could even be an instruction to the large language model to generate its response in a different language.

Foundation models can be tuned to further improve the performance of the model for a specific task. We can get significant improvements in the model performance by tuning the model on a modest number of example data. The new fine-tuned model is effectively a combination of the newly learned parameters and the original model. Since we are not training the model from scratch, it is both data- and cost-effective.

Another benefit of foundation models is that they are more resilient to noise, or errors in the data, than traditional AI models. This resilience is crucial since real-world data is often quite noisy.

Foundation models ultimately help make GenAI more powerful, flexible, and efficient than other AI models.