There’s certainly a lot of excitement about Large Language Models (LLMs) and the opportunities they provide for transforming how we interact with technology. LLMs are an evolving technology with a wide range of use cases and they are part of the wider field of Generative AI (AI that can generate new content). An LLM typically consists of a neural network trained on vast quantities (think billions!) of text using some kind of self-supervised or reinforcement learning. The result is a system that can understand plain language as well as generate it. LLMs can recognize, summarize, translate, predict and generate text. ChatGPT is an example of a large language model and there are many others.
So how can this technology be applied to the day-to-day operations of a business? There are a number of approaches to consider.
- Use the LLM as a standalone tool
In this scenario, users interact directly with the LLM to boost productivity. For example, your employees could ask chatGPT research questions or ask it to summarise a document they need to digest. This is a great approach for initially learning about the technology and to get some quick wins when dealing with large blocks of text without the need for any IT investment. A watchout with this approach is that any data you submit to the LLM may be available to third parties.
- Use the LLM to create prompts
LLMs work by interpreting and processing prompts from the user. Prompt engineering has become a discipline in and of itself: mastering the perfect prompt to provide to the machine. In what may initially seem like a highly paradoxical use case, the LLM can itself be used to generate better prompts. For example a user can request that the LLM expand or augment a prompt by adding in details.
This same approach can be used in image generation, a related field of Generative AI. Perhaps you’ve seen some of the entertaining images generated by asking a machine to paint someone’s pet in the style of Van Gogh or some other obscure combination of artistic styles? LLMs can generate the prompts that make these images, augmenting a user’s initial textual description. The impact is quick customization of images but there remains a lot of unanswered questions about the copyright implications of generated images.
- Embed LLM into an app
Embedding LLMs into software applications is a more advanced technique and provides an opportunity to surface LLM output alongside other operational outputs. In this approach, a piece of software makes API calls to the LLM and surfaces the results for the user inside of an interface they are already using. The advantage over option (1) is that the user does not have to leave their normal environment and can access the LLM where they are. This approach also can help improve privacy and security as API calls can be logged. Sophisticated workflows can be setup where controls, monitoring and gating are put in place around the API calls. A real world example of this approach would is including LLM search capabilities within a web browser or word processor.
- Embedding LLMs into workflows
An iteration on the previous example, this approach involves building calls to the LLM into an application workflow. Rather than simply exposing the LLM API inside of an existing user interface, here we are talking about truly integrating the LLM into the functionality of the application. An example would be integrating LLMs into productivity or collaboration software, where suggestions from the software are being augmented automatically by the LLM, perhaps even without the end-user’s knowledge. Another common use case is using LLMs to generate website content within a content management system.
- Fine tuning an LLM
The final design approach to consider involves taking an existing LLM and fine tuning it with a business’ own documents, content, metadata or other data. Unlike generating better prompts or providing different ways to access the LLM, this approach alters and updates the underlying model. There is an incredible opportunity to adapt existing LLMs for specialised use cases and tasks using your own organisation’s data and semantic context. For example a retailer could create an LLM that is well-versed in all of its products and their individual specifications and sales histories, as well as all of its customers, allowing team members to make detailed queries using plain language about product performance. Or a financial services business could fine-tune an LLM so that it understands the inner workings and gearing of all of its lending products so it can be used to assist with credit approvals. Another use case is to make chatbots more sophisticated and knowledgeable about a business’ operations.
In conclusion, there are many exciting ways to use LLMs and this technology will only continue to evolve. Look out for natural language processing in many of the tools and systems you use.