Maximizing AI ROI: Why bigger models aren't always better
NOV. 21, 2024
Bigger artificial intelligence (AI) models don't always mean better ROI. Tailored, efficient small models can do more for your business at lower cost and help optimize your AI investment.
It's been almost two years since OpenAI launched ChatGPT, and in that time, many leaders have started placing AI at the center of their business strategy. And while this technology has the potential to revolutionize how businesses operate, the big AI models that characterize it are proving to have limitations for solving specific business problems.
The thing is, when it comes to AI in business contexts, less can often be more. Tailored and efficient small models can do more for your business at a fraction of the cost and help optimize your AI ROI.
The big model paradox
Large, complex AI models—also known as large language models or LLMs—can engage in human-like conversations, generate creative content, and process vast amounts of data in seconds. However, when it comes to specific business use cases, adopting a large AI model can be like choosing a sledgehammer to hit a single nail into a wall.
Big models and their promises
Big models promise groundbreaking capabilities but often underdeliver in specific business contexts. They're jacks of all trades, doing a good—but not great—job at most tasks. For businesses to be more effective, they require tailored models explicitly suited to address particular tasks.
Let’s use a financial services company as an example. Say the company wanted to create a chatbot that answered questions about specific products, like their credit card offerings. While a big model could do an adequate job at this, it would be expensive to run for just this task and harder to train. A smaller, more tailored model could be set up with the right information and made into a specialist (of sorts) on those products.
The hidden costs: Computation, maintenance, and expertise
Big models are riddled with hidden costs. They demand enormous computational resources, often requiring specialized hardware and substantial energy consumption. Ongoing maintenance is a constant challenge, requiring regular updates and fine-tuning.
Moreover, leveraging these complex systems and making them as effective as possible requires specialized expertise. Businesses need to hire data scientists and engineers focusing on AI—a costly proposition given that expertise in this area is still in its nascent stages.
The diminishing returns of complexity
Increasingly complex models often deliver diminishing returns in performance, especially for specific business tasks. As teams add more parameters or data, they may see marginal gains while costs keep rising. In a classic case of the law of diminishing returns, tech leaders often pay more for fewer significant advancements, ultimately reducing their AI ROI.
The case for small models
AI for business doesn’t just exist in the form of big models. Instead, businesses can leverage smaller, simpler AI models focused on capabilities that matter.
Defining small models: Focused, efficient, and task-specific
When we talk about small models, we mean that they’re purposefully designed to be focused, efficient, and task-specific. They’re machine learning algorithms trained on smaller and more specific troves of data, making them more agile and accurate.
These models—sometimes called small language models or SLMs—address particular business challenges, are optimized for specific data types and workflows, and deliver maximum efficiency in their designated tasks.
Some examples of where small models do their best work include locally hosted systems (e.g., an internal wiki), chatbots, and mobile applications.
Advantages: speed, cost-effectiveness, and interpretability
Tailored models excel in speed of delivery, cost-effectiveness, and interpretability—crucial factors for real-world business deployment. They require less computational power, translating to lower infrastructure costs and reduced energy consumption.
“The focused nature of small models tends to make them more accurate and also means that they can be fine-tuned faster to the nuances of particular datasets or applications," says Fred Souza, JavaScript Engineer at Lumenalta. "It also makes them easier to deploy, maintain, and update.”
At Lumenalta, we’ve been testing the efficacy of different LLM sizes for an internal workplace AI. When comparing Llama 3.1 8b vs. Llama 3.1 405b the same simple prompt rules, the output results were virtually the same—but the larger model cost up to two times more than what we would pay for the smaller model.
Interpretability is another key advantage. Tailored models often provide clearer insights into their decision-making processes, a critical factor in industries like financial services where explainability is paramount for regulatory compliance or building user trust.
Successful small model implementations
While they may not take up the spotlight like big models do, there are many examples of simpler models streamlining operations for businesses in different industries. For instance, retail giants have implemented tailored demand forecasting and inventory management models, significantly reducing overstock and stockouts.
Manufacturing firms leverage specialized quality control AI models that provide real-time data and dramatically improve defect detection rates. In the financial sector, companies are refining their investment advice with tailored models that cost much less than LLMs like GPT-4.
Aligning AI strategy with business outcomes
Aligning your AI strategy with specific business objectives is key to maximizing AI ROI. Here’s what that looks like in practice:
Ensure you have clearly defined business objectives
Successful AI implementation starts with clearly defined business objectives, rather than technological capabilities. Starting with your goals ensures that your AI solutions are properly tailored to your needs, avoiding unnecessary complexity. By building models to the scope of business value, you ensure they are properly tailored to your needs and deliver maximum ROI.
Match model complexity to problem complexity
Align your model’s complexity to the problem at hand. For many business challenges, a focused, efficient model will outperform a more complex one. Don’t waste time and resources turning the model into a bigger solution than necessary. This approach not only saves resources but often leads to increased accuracy, efficiency, and—depending on the purpose of the model—customer trust.
Let your data strategy inform your model selection
A robust data strategy determines the correct model size and complexity. For starters, you should be clear about data’s role in your organization and its quality, quantity, and relevance. Your strategy will also inform the data infrastructure, which will, in turn, determine the type of models you can build and adopt.
Solving the implementation issue
Implementing tailored AI solutions has its own set of challenges, but the right approach can address these.
Overcoming the
To disrupt the flawed belief that bigger models are always better, you must educate your teams and demonstrate the value of targeted solutions. Showcase early wins with tailored models to build confidence in this approach throughout your organization. Highlight the tangible benefits in terms of speed, cost, and effectiveness.
Building internal expertise for long-term success
Invest in training programs that empower your team to develop, deploy, and maintain efficient AI solutions. Establishing a strong internal capability will allow for greater flexibility and faster iteration on AI projects, reducing dependency on external vendors.
Streamlining AI project management
Adopting agile methodologies in AI project management enables iterative improvement and faster time-to-value. By continuously refining your AI models, you can keep them aligned with evolving business needs.
The human factor in AI success
The success of small AI models isn’t just about the technology—there’s also a human element at play.
User adoption and trust in simpler models
User adoption and trust are critical for AI success. Because of their efficacy and higher accuracy, simpler, more interpretable models often garner higher trust from employees. This makes them more likely to be adopted, increasing the model's impact on the organization.
The importance of explainability in AI decisions
Explainable AI is becoming increasingly important, not just for building trust but also for meeting regulatory requirements in many industries. Tailored models make it easier to understand and explain the factors contributing to outputs compared to their more complex counterparts.
A culture of AI literacy
As AI grows in usage and popularity, fostering a culture of AI literacy across your organization is essential. When employees understand AI's capabilities and limitations, they can leverage these tools effectively and contribute to their ongoing improvement. The more your employees understand how to use AI, the more they'll adopt AI into their day-to-day work.
Future-proofing your AI strategy
The AI landscape is evolving rapidly, requiring strategies to adapt to new developments. To stay competitive without chasing every new trend, balance innovation with pragmatism. Develop a flexible, scalable AI infrastructure that easily integrates new technologies as they become relevant to your business needs.
Focusing on tailored, efficient AI solutions aligned with your business objectives will allow you to maximize your AI ROI. These focused models deliver better results and offer cost savings, improved interpretability, and easier implementation. In the world of AI, sometimes less truly is more.
