top of page
Writer's pictureMackenzie Robey

Would a tailored LLM be useful for my organization? What are the main drawbacks?

Large Language Models (LLMs), like GPT4, Claude-2, and Llama2 (in the form of apps like ChatGPT, Perplexity, Bard/Gemini , and others) have gotten a lot of attention in the business world due to their potential to enhance and simplify various operations. LLMs are a type of AI model that can produce text and code responses based on an input question and potential follow-up questions/context from the user (or query). These models were trained on massive training sets, where they analyzed text with diverse language patterns in a mathematical way to be able to answer questions in a human-like form.


While it seems they are creating something out of thin air, they are truly just answering an extraordinarily difficult math problem based on statistical associations within the training data. In other words, they are generating answers based on analyzing the data provided. The common phrase "you are what you eat" could be applied to LLMs as "LLMs are what they analyze" - they can only produce quality information if they were trained on quality information.


Although they are helpful in providing answers to generic questions, a lot of businesses have questions that are very specific to their internal data sets and systems. With the development of Retrieval Augmented Generation, it is now possible to tailor the implementation of an LLM to be able to reference pertinent company information without having to retrain or fine-tune as illustrated below.




Tailoring the implementation of an LLM for your business

The development and advancements of Retrieval Augmented Generation has caused a massive expansion of LLM applications. Instead of digging through disconnected legacy data systems, you could ask your organization's LLM for specific, internal information and get it immediately. Instead of trying to find meeting notes to get that quarter's sales numbers, your organization's LLM could accurately provide them in seconds.


LLMs are powerful tools to greatly reduce the time required for finding and analyzing internal information - however, they can also be used for organizational-specific content creation as well. Some examples of LLMs used in cases where information is generated are:


  1. Marketing - They can generate on-brand, meaningful, and custom sales content to speak to your customer.

  2. Customer Service - Customer service advocates can greatly benefit from custom LLMs as they can help answer specific customer service requests, access internal sales and shipping data, and write tailored responses to help your customers get their questions and concerns handled in a fraction of the time.

  3. Legal Services - LLMs trained on legal records can help generate specific and meticulous drafts of documents such as wills, trusts, and contracts. They can also be provided documents and asked to check for errors.

  4. Transportation - LLMs can analyze traffic patterns, route plans, and safety data to help make informed decisions.

  5. Retail and eCommerce - LLMs can help customers quickly answer product questions. They can also help make recommendations to customers based on budget and design preferences.

  6. Consumer Goods - LLMs can help analyze customer information and market trends to provide keen insights during the development and deployment of consumer goods.


This is not a comprehensive list by any means - LLMs are powerful and adaptable. An LLM could benefit most organizations that have a large amount of word and number-based organizational data to make interacting and accessing it efficient and seamless.


If you believe that implementing a tailored LLM in your business may be a good fit, these four questions will help you determine what that implementation should look like.


4 key questions to consider before researching LLMs


How will your organization interact with the model?

Are you looking for a chatbot for customers to help answer questions in a more personalized and nuanced way compared to traditional chatbots? Or a tool to acquire internal company data and do automated analysis? Questions about who interacts with the model, how many people, and their expertise on the information should be considered.


Where is your company data stored?

To be able to provide accurate responses, accurate and domain-specific data needs to be provided. Large organizations, especially those that have been around a while, can have data stored in multiple antiquated locations in different formats. Obtaining as much relevant information as possible to provide to the LLM will lead to the best results. Also, models will need to be updated regularly where the frequency depends on the nature of the information. Ensuring your organization's data can be easily accessed to allow for quick retraining will help make the model's accuracy and usage sustainable.


How will the data need to be preprocessed to provide it to the model?

Different LLMs require your organization's data to be preprocessed differently. Preprocessing can be a nightmare depending on the mode in which it is conducted so understanding what data types your model can analyze, what form your current data is in, and how it needs to be changed to work with the desired model is paramount.


Do you have the internal expertise to create and manage your LLM?

It is possible to train your model yourself or reach out to a company that specializes in ML services to help you do so. If you have in-house expertise, there are thousands of pre-trained models to choose from, and HuggingFace Model Hub, TensorFlow, and PyTorch are the go-to locations to acquire a pre-trained model. With thousands to choose from, it is recommended to try out multiple models based on your computing needs, required accuracy, and input data format when selecting one.



Drawbacks of customizing LLM integration

Resource Intensiveness: The development and training of LLMs require a substantial investment of time and resources. This can pose a challenge for organizations with limited processing capabilities or tight timelines, impacting their ability to deploy tailored LLMs promptly.


Domain Limitation: Tailored LLMs are less versatile outside their trained domain, which means they may not perform as effectively when faced with tasks beyond their specific focus. This limitation is evident in industries where a broad range of subject matter is being utilized, such as in legal services and healthcare. Here is a fascinating article about legal LLMs potential to generate false information, or "hallucinate".

.

Ethical Considerations: LLMs can potentially absorb biases from their training data, leading to negative ethical implications. Organizations must carefully address this issue to avoid unintended repercussions, especially in sectors where unbiased decision-making is critical, including finance and healthcare. This article provides a deep-dive of ethical considerations in healthcare, including the potential of LLMs to continue hidden biases based on race, gender, and income that could lead to unequal patient outcomes.


Training Data Risks: Incomplete or inaccurate training data for tailored LLMs can pose significant risks for organizations, potentially leading to false market trend predictions or security issues. Also, the server location of the LLM and security around who has access to the training data needs to be monitored. Though all industries deal with matters of security, retail, healthcare, and cybersecurity are industries where data accuracy and security are paramount.


If you're interested in learning more, schedule a call with us and we would love to talk with you about the potential of LLMs within your organization.


Sources:




Comments


bottom of page