To Build or not to Build - is that the Question?

Cracking the ‘Buy’ or ‘Build’ conundrum in Gen AI – a PoV on when to & when not to use custom LLMs.

The Generative AI is the next revolution in technology and has taken the world by storm ever since the release of ChatGPT. There is a potential for it to be infused into most businesses – whether the use cases are table stakes or strategic.?

Given the multitude of players in the space and the multi-dimensional aspect of the usage of a technology such as this, we are attempting to start a discussion on which models fit best.?

What are the patterns that emerge in how one could apply foundation models??

There are three key patterns in which you can engage with foundation models – “Buy, Boost, and Build”.??

• “Buy” is simply licensing an existing model & leveraging prompting to get the best results.?
• “Boost” is licensing existing models & enhancing them via RAG or fine-tuning.??
• “Build” is training a custom model from the ground up.?

The ‘Buy’ or ‘Build’ Dilemma?

Gen AI is disrupting businesses as we know them, and organizations are racing towards becoming AI-first organizations. Democratizing Gen AI can be the difference between a successful organization, and an unsuccessful one.??

While its potential is clear, the path to harnessing the power of Gen AI involves intricacies that require various careful considerations. IT leaders are required to act quickly and make a multitude of critical strategic decisions. After identifying the “lighthouse” use cases for piloting Gen AI, the next important question is whether to ‘Buy’ an existing model or ‘Build’ a custom model from scratch. ‘Boost’ is considered a special case of ‘Buy’ for the scope of this article.?

“To build or not to build – that is the question.”?

Factors to Consider?

At this critical crossroads, the decision is beyond just a matter of technology. The choice one makes will have lasting implications across the organization – shaping its pathway, preserving its autonomy, and optimizing resource consumption while maximizing impact. ?

When weighing the pros and cons of the ‘Buy’ vs. ‘Build’ decision for foundation models, 3 key factors emerge:?

1. Privacy, Ownership, and Control: ‘Build’ will provide more privacy & control due to model ownership, autonomy, increased flexibility of custom models.?

1. Resource Requirement: Training from scratch tends to cost more, consume additional compute resources, and entail higher carbon emissions. It even requires specialized skills and vast amounts of quality data.?

1. Value Realization: Finally, the decision boils to maximizing value realization for Gen AI initiatives.?

Let's delve into specific scenarios to understand these trade-offs in greater detail.?

01. Privacy, Ownership, and Control?

When to ‘Buy’?

• When you only need the general-purpose capabilities from LLMs and there is little to no need for task-specific customization.?
• When in the prototyping/exploration phase, ‘Buy’ is a more pragmatic choice – which also allows for benefiting from collaboration and community improvements that open-source models offer.?

When to ‘Build’?

• When there is a need for domain-specific models offering higher accuracy and improved explainability – especially where unique requirements, such as longer sequence length, specific content filters, brand voice, specific language support, dialect-specific understanding, etc. are not met by existing models.?
• When there is a necessity to handle sensitive or proprietary data with utmost confidentiality, protecting against exposure, particularly when privacy-enhancing techniques are insufficient.??
• When existing models fall short in meeting performance and other critical criteria, despite efforts to enhance them through transfer learning, RAG, prompt engineering, etc., there arises a need for greater control over the model.??
• For full autonomy over training datasets used for the pre-training, which directly impacts model quality, bias, and toxicity issues.?
• When the model is a core part of your business strategy and technological moat/ competitive advantage, building custom models becomes imperative.?

02. Resource Requirement?

When to ‘Buy’?

• When organizations have a low risk-appetite, building from scratch can be expensive and extensively risky if not done with utmost care. It also helps to reduce lock-in, as investments in building a model that does not deliver as expected can lead to burdensome sunk costs.?
• When there are resource constraints around cost, carbon, and/or compute, building may not be the ideal choice. Buy is especially more suitable in this scenario as AI licensing costs tend to be lower whereas training LLMs has substantially high requirements of specialized hardware (e.g., ChatGPT was trained on 25000+ GPUs in a matter of days exploiting the parallelism provided by GPUs, whereas on a single GPU, it would take 350+ years).?
• When the right talent with specialized skills required for ‘Build’ is not available and it is difficult/ costly to find new talent or upskill/cross-skill existing ones.?
• When you have limited data available for customization, it is better to opt for boosting existing models with RAG, fine-tuning etc.?

When to ‘Build’?

• When the right talent, technology, and resources, such as funding, compute, quality data, time, etc., required for development are readily available.?

03. Value Realization?

The decision to ‘Buy’ or ‘Build’ is fundamentally an evaluation of impact to maximize value realization – where technology stands at the center of business transformation. Value realization being subjective, there is no direct way to recommend whether to ‘Buy’ or ‘Build’. While there is no uniform way to calculate it, AI impact could be visible to an organization in various form factors:?

First is Value Creation via…?

• Faster value realization?
• Improved quality & performance?
• Operational efficiency gain?
• Process augmentation?
• Driving innovation via new product or service opportunities?
• Enriched experiences (of both customers and employees)?

Then comes Tactical Fortification including…?

• Enhanced security and privacy?
• Long-term talent optimization?
• Improved sustainability?
• Reduced risk & increased transparency?
• Sustainable competitive advantage or technological moat creation?
• Long-term goodwill acquisition?

Eventually leading to Financial Gains in the form of…?

• Lower bottom line & increased profitability?

• New revenue channels & business models?

The benefits of AI projects may not be realized instantly and may take months or years to manifest. Return on investment (ROI) evaluations are critical to ensure that AI initiatives deliver tangible value at speed, justify the costs & risks associated to them, and continue to contribute to the long-term success & well-being of the organization as well as society.?

Unlock a Frontier of Possibilities?

Whether you choose to steer your own ship, or catapult the existing, generative AI remains a frontier of possibilities. While unlocking this frontier, it is crucial to embed responsible AI practices addressing ethical considerations, ensuring transparency, and maintaining accountability.?

Gen AI is transitioning from an epic differentiation to a point of parity. There are clear signals of change required in the way we do business, and merely reacting to these changes is insufficient. Now is the time to ride this wave of change and invest in responsible generative AI initiatives, to super-charge our businesses and emerge as leaders in this dynamic landscape.?

?

Look Ahead?

Having navigated the ‘Buy’ or ‘Build’ dilemma in the world of generative AI, our focus now shifts to a more fundamental debate – when to use classical AI and generative AI. In our upcoming blog, we will delve into these 2 distinct realms along with their pros & cons - explore the future of classical AI in the Gen AI era. Will classical models fade away, or can they coexist harmoniously with the innovative power of generative AI? Join us as we unravel the dynamics shaping the future of Artificial intelligence.