With many organisations already leveraging generative AI (GenAI) to drive business decisions, draft customer and client interactions, and analyse vast sets of big data, harnessing the power of a large language model (LLM) is the natural next step. An LLM brings a host of benefits and efficiency gains with them - creative content, solutions to complex problems, and novel product and service suggestions.
The key question, however, is what type of LLM would best suit your overarching business goals and operations. Different LLMs have unique advantages and disadvantages - and you don’t want to find this out after already making the necessary cost and time investments.
So, should you opt for an open-source LLM, a proprietary LLM, or go for a completely DIY approach? Let’s examine LLMs more closely below.
What is the difference between open-source and proprietary (closed-source) LLMs?
Firstly, what is an open-source LLM? Simply put, an open-source LLM is built using code, model weights, and underlying architecture that are accessible to the public—although in some instances (like with Llama), the training dataset is not. This means that developers and researchers can access the code and model weights to collaborate and improve the model’s capabilities. An open-source LLM is publicly available and offers full transparency regarding the code, architecture, and parameters used to build it, allowing businesses to inspect it for capabilities, limitations, and potential biases.
A closed-source or proprietary LLM is developed and owned by a company, using private training data and resources. The data used is seldom private—it is often sourced from public sources. However, you don't get to see what that training data looks like, so it's private from an access perspective but not exclusively from private sources. Unlike open-source LLMs, a proprietary LLM is not publicly available and is accessible via the company, usually through a monthly subscription payment structure. While its underlying code and architecture might not be freely available for assessment, a proprietary LLM is generally trained on massive datasets, offering robust capabilities.
Which LLM option is the better choice for organisations?
Both open-source and proprietary LLMs offer unique advantages. A proprietary LLM is generally trained on large datasets; however, the source and type of data used are not disclosed. Accountability for the LLM’s performance rests solely with the corporate entity responsible for building and training it. It’s also in the best interest of the founding company to continuously update and improve its model’s capabilities to sustain revenue and public interest.
But, these models are fixed - the training data used to train them is undisclosed and not open to customisation. Larger LLM models like GPT-4o may have colossal processing power well-suited for generalist operations, but lack the niche specialisation that comes from unique training data - your business’s data! Subscription and licensing fees can be hefty to maintain.
An open-source LLM is free to access and use. Their transparency means you can carefully inspect their structure and, in some instances, their training data (though in cases like Llama and DBRX, the training data is not visible) before deciding to invest in a model. Open-source LLMs allow for customisation and fine-tuning for specific use cases and niche operations.
This is because an open-source LLM allows you to add your data to customise its training and outputs, giving it the unique context necessary to generate highly tailored responses to your submitted prompts and queries. An open-source LLM also has foundational training already, meaning you don’t need to train it from the ground up.
An open-source model can allow for greater customisation and fine-tuning based on domain and use-case-specific data for niche outputs unique to your business needs. However, whether it is more cost-effective compared to a proprietary LLM depends on the specific use case.
However, an open-source LLM doesn’t have the same final accountability for its performance that a proprietary model does or the financial backing and support of a parent company or organisation. That being said, open-source LLMs are still a fantastic, cost-effective solution for businesses looking to build more tailored, case-specific solutions.
Should you build your own LLM?
Only if you’re looking to test the limits of your patience and budget. Building your own LLM is generally not a feasible option for most organisations due to several significant challenges. The process is extremely expensive and time-consuming, requiring vast amounts of general training data and substantial computational resources. Additionally, you need to invest in storage space for these resources and hire a specialised team of data scientists, machine learning engineers, and NLP experts to develop and train the LLM. Access to GPUs, which is a critical resource for training LLMs, represents one of the biggest constraints and costs in this endeavour.
The time required to train the model to a usable state can span months, adding to the impracticality. Given these constraints, it's advisable to consider fine-tuning an existing LLM for your specific use case rather than attempting to build one from scratch.
Think carefully before you invest
The LLM you choose to use for your GenAI operations across all departments and use cases will directly affect your organisation’s performance. It’s an investment on all fronts - time, team and budget. How successful your efforts are will depend greatly on the model you decide is best for your operations, making it all the more vital to ensure you are investing in the right solution.
Carefully consider your business size, unique needs, current challenges that could be solved through GenAI capabilities and your long-term key objectives. It’s not a decision you should come to quickly but carefully thought out, with consultation and advice from external experts if need be.
