What is an LLM?

LLM’s learn patterns from giant datasets of media and produce multi-modal content.

An LLM is a kind of neural network that is trained on a massive dataset of human-created media (text, audio, images, and more). They are designed so that when you give it an input (otherwise known as a prompt), it will generate an output based on your prompt and the patterns it learned from its training dataset.

LLM’s can be used to create queries (SQL,KQL), code (Python scripts, Java), images and videos that can capture your imagination, audio that sounds like celebrities, and text that can be either creative or research-focused in nature. Keep in mind that while the text it produces is human-sounding, it is not entirely accurate. LLM’s merely predict the next word based on three factors: the previous words they have generated in the interaction, the prompt, and its training data.

Fine Tuning

Most LLM’s were created to be general-purpose, and based on their giant dataset and the Deep Learning techniques that trained them, they have a lot of connected information they rely upon. However, when you need an LLM to be very specific, you can give it data for fine tuning, focusing its informational universe on the tasks you want it to execute. Good examples of this are a document containing company’s knowledge base that helps them execute customer service requests. By being able to have data that is directly related to what the company does, and focusing all queries on that data, the LLM is able to extract more information in a faster and more accurate way.

Agentic Workflows

A fascinating application of LLM’s are agentic workflows. Usually, people go into an LLM’s interface and type in what they want and get an output. What an agentic workflow does is it divides a single prompt into multiple input-output pairs: input #1’s output becomes prompt #2’s input and so on until you have completed all the steps related to the task. It is a way of ensuring more thought and computational energy goes into producing the overall result. Think of it like this: you’re using a single prompt, converting it into 2 (or more) prompts, and creating a single, more refined output. This can be done on the backend in a programming-defined way to make things far more interesting. The game with this is just beginning. Add fine-tuning data that maps on to your goals and you’re off to the races.

For example, when you tell the LLM: plan a project for writing a new application. It will divide that prompt into multiple prompts and steps that create the final output, like: 1) “state the goal of creating a project for a new application”, 2) “outline all the necessary steps that are necessary to complete the project from start to finish”, and 3) “then generate a succinct outline of all the steps”. I’m sure you can get more creative than that. Each time, the outputs can be used to give the model additional context. Now imagine creating plans for: creating a new material that replaces plastic, designing a GPU, and picking AI companies to invest in.

Giving Context

A cool idea I’ve always thought about is teaching an LLM to become a wizard, giving it fine-tuning data that will make it respond as a wizard and infuse additional customized context into your responses. You can give the wizard .txt files that show how the wizard has operated in the past and the things is should know (like where the medicinal herbs are in the forest). The LLM will draw from its knowledge of the world to ‘understand’ what a wizard is through a contextual search of its training data and the fine-tuning data you have provided. You could add agentic workflows in the code in order to get more powerful and refined answers.

Conclusion

Large Language Models are a new technology that are revolutionizing the way data is created. Companies like IBM are getting after it to ensure that business are maximizing the value of this new and dynamic tech stack. I’m excited to continue writing about this topic. Feel free to leave any questions or additional topics you’d like me to research in the comments. This technology is just the beginning of a new revolution. Stay tuned, stay safe, and make sure you use it right.