What do LLMs mean to your business?

Imagine a world without Google Translate, Siri, or autocomplete. That would feel like a world a long time ago in a galaxy far, far away…

ChatGPT was not the first LLM-based application or even chat model. However, ChatGPT made LLM-powered applications widely accessible, starting new conversations on how LLMs can be used in real-world business applications. Prior to these advancements, producing a model on text data required years of experience with NLP, statistics, and programming. Leveraging that model also required significant domain knowledge and was only applicable to narrowly-defined use cases.

Now, LLMs are available both as open-source models and SaaS offerings that provide significantly better quality than traditional NLP. For example, tasks like transcription can be done with a few lines of code, deliver better quality than commercial services at a lower cost, and can be completed in a fraction of the time. Let’s see a few applications of LLMs:

Now that you’ve seen a few examples of applications using LLMs, let’s discuss a few factors to consider when using them.

Prompt engineering is a common technique to elicit model responses, aligned with your use case. Here, the description of the task is embedded in the input. For example, in question/answer use cases, a prompt may be:

Spoiler Alert: You would expect the answer to be “Darth Vader” or “Anakin Skywalker”.

Throughout the rest of this blog, when we use the term “fine-tune” we are referring to the more general process of using your data to update an existing model.

For example, suppose you’re a medical company trying to analyze insurance claims. In that case, you can bet that medical insurance data wasn’t included in the main training corpus (or else there would be serious GDPR and/or HIPAA concerns). As a result, many companies will take these pre-trained LLMs as foundation models and then fine-tune them with their own data. Pre-training these foundation models from scratch is prohibitively expensive in most cases, as it requires highly-specialized knowledge and huge investments in data gathering, training computation resources, and validation. Based on the task, data, and resources, your LLM-based approach will vary.

Don’t be mistaken that LLMs are a general purpose solution for your business use case. For example, if your company is predicting stock prices, you could absolutely use LLMs as one component of your model, say to incorporate textual information (e.g. decoded messages intercepted from the Rebels). However, you’ll still want to combine it with all of your numeric data (e.g. from financial models, activity in local Trade Federation routes, etc.) for a comprehensive solution. At least we would hope that you wouldn’t predict stock prices solely based on Rebel communiques.

If you use an off-the-shelf LLM with no modifications, you have limited competitive advantage. While ChatGPT is a free-to-use application through its web interface for non-commercial purposes, you will need to pay for anything beyond that. OpenAI provides hosted instances of models that you can fine-tune with your data, however, currently, you will need to send that data to their servers and can only query those models; you won’t have direct access to them or full insight into how the underlying model was trained. If you’re a well resourced Republic, you have the luxury to build your own LLM from scratch. But if you are a scrappy Rebellion, you will have to use an existing take an off-the-shelf model and update it with your data. Let’s contrast these two approaches.

Now the real question is whether you do the fine-tuning in-house or rely on an external service. Depending on the application, you may wish instead to own the model entirely and use an open-source pre-trained foundation model. This would involve downloading and storing it to allow different parts of your business to fine-tune this foundation model for specific tasks. Therefore, as new open alternatives become available, you can remain in control of your NLP applications. In many cases, we recommend doing it in-house so you have full control over the model and data, and retain your competitive edge.

Typically the more data an LLM is trained on, the better its performance. However, many business cases require only a fraction of the power these massive models can produce and still yield high business impact. You need to consider an acceptable trade-off in terms of compute vs cost as larger models are more expensive both for model training and making predictions. Models like GPT-4 are so large because they’re general purpose, but most use cases in industry are on more narrow applications only needing smaller models.

“Patience you must have, my young Padawan.” Hopefully, at this point, you realize that LLMs are far more than just chatbots and that, as with all technological advances, you will need to decide which direction to take if you want to harness their power. For some, a hosted proprietary product will suffice where ownership of models and custom deployment are not critical issues. For others, a DIY approach will be preferred, if not needed, where an open-source, pre-trained LLM can be molded to suit your use cases using your data and domain expertise. Regardless, the decisive ingredients that determine your success will always be the technical expertise, data, and computational resources you have at your disposal.

May the Fourth be with you!

Related posts:

i ii Much the same idea as last time to convince anything against her. t papa. you’ve ever made me dfrom that he is already rock in a different variations body, high Dreams has always lost his girl…

People are always happy working from their zones of perfections. Most people fear to work at times and in environments that they know they can’t thrive well. As it is always referenced that when…

You took an oath to uphold the Constitution of the United States. I firmly believe that we are in a constitutional crisis now. Our executive branch is in meltdown. It is now clear to anyone with eyes…