From the course: Introduction to Large Language Models (LLMs) and Prompt Engineering by Pearson

Applications of LLMs

From the course: Introduction to Large Language Models (LLMs) and Prompt Engineering by Pearson

Start my 1-month free trial Buy for my team

Applications of LLMs

“

And when we think about applying LLMs, we generally do this in three ways. Number one is we can encode text into vectors with little to no fine-tuning. This is going to be how we use LLMs in our next lesson when we build a semantic search system using BERT and OpenAI as well. Number two is fine-tuning an LLM to perform a very specific task with custom data. This is along the lines of using transfer learning, where we'll take the pre-training of an LLM and rely on it to learn a new task that is more nuanced. The third way we'll use LLMs is arguably the simplest way, which is just to ask an LLM to solve a task. Like you would go to chat.openai.com and type in something to chat at GPT and get an answer back. That's number three. We're going to call that prompting more formally later on. But the idea is we simply ask an LLM to solve a task, whether it's open source or closed source. No fine-tuning, no low-level vectorization of data, just simply asking the model to do something with specific instructions, frameworks, examples, validation, and getting an answer are back that we can actually use in line with another workflow. This is going to end up being the three ways we generally are going to be using LLMs together. Of course, there are challenges. I've already named a few. But to rename a few of them, LLMs are generally larger than classic machine learning models like RNNs and LSTMs. And they can be more difficult to manage, at least in production, without proper ML ops, or as we will generally call it in our time, LLM Ops. But ML Ops and LLM Ops aren't so different, but there are some nuances between them, which we'll likely see in later lessons. And choosing which LLM to use, as I mentioned, will depend on the task that you're trying to solve and the knowledge the LLM needs to have to solve that task. And that will tell us whether we can simply prompt it for an answer, fine tune it, or encode the text and use it in a different pipeline. And the third challenge, not the only three challenges, but a third challenge that I'll reiterate is LLMs are biased. And I'm using the term biased here a little bit stricter than how it might be used in general media. When I say biased output, what I literally mean is an LLM is only able to solve a task based on data it has already seen and whatever you gave it in its input window, that prompt that you gave it. Those are the only two things at a high level that an LLM has access to. So an LLM will be biased based on information it has seen being stale or wrong or harmful. And we'll combine that with what you asked it to do. And that's how you can bias, i.e. alter the output of an LLM to be malicious or harmful or whatever a human so chooses to do. So we will be ethical human beings and try not to do that. And we'll also see some more rigorous ways of preventing and identifying it as we see applications of LLMs.

Contents