Scale generative AI with flexible model selection

This blog series explores enterprise generative AI (gen AI) for business and technology leaders. It provides a simple framework and guiding principles for your transformational artificial intelligence (AI) journey. In a previous blog, we discussed IBM’s differentiated approach to delivering an enterprise-grade model. In this blog, we explore why choosing a foundational model is important and how it can help companies scale their AI generations with confidence.

Why is model selection important?

In the dynamic world of the AI ​​generation, a one-size-fits-all approach is inadequate. As companies seek to harness the power of AI, they will need a variety of model choices to:

• Accelerating Innovation: A diverse palette of models not only fosters innovation by providing distinct strengths to solve different problems, but also enables teams to adapt to changing business requirements and customer expectations.
• Customization for Competitive Advantage: A variety of models allows businesses to gain a competitive advantage by customizing AI applications to niche requirements. Gen AI can be fine-tuned for specific tasks, whether it’s a question-answering chat application or writing code to generate quick summaries.
• Shorten time to market: In today’s fast-paced business environment, time is of the essence. A diverse portfolio of models accelerates the development process, allowing companies to quickly adopt AI-based products. This is especially important in Gen AI, where access to the latest innovations provides a significant competitive advantage.
• Respond flexibly to change: Market conditions and business strategies are constantly evolving. A variety of model choices allows companies to transition quickly and effectively. Having access to a variety of options allows you to remain agile and resilient by quickly adapting to new trends or strategic changes as they arise.
• Cost optimization across use cases: Different models have different impact on costs. Having access to a variety of models allows companies to choose the most cost-effective option for each application. Some tasks may require the precision of more expensive models, while others can be addressed with cheaper alternatives without sacrificing quality. For example, in customer care, throughput and latency may be more important than accuracy, but in resources and development, accuracy may be more important.
• Risk Mitigation: Relying on a single model or making limited choices can be risky. A diverse portfolio of models helps mitigate concentration risk, ensuring companies remain resilient to the shortcomings or failures of one particular approach. This strategy allows for risk diversification and provides alternative solutions if problems arise.
• Compliance:The regulatory environment for AI continues to evolve with ethical considerations at the forefront. Different models can have different impacts on fairness, privacy, and compliance. A wide range of choices allows companies to navigate this complex landscape and choose a model that meets their legal and ethical standards.

Choose the right AI model

Now that you understand the importance of model selection, how do you address the problem of selection overload when choosing the right model for a specific use case? We can break this complex problem down into a series of simple steps you can apply today.

1. Identify clear use cases: Determine the specific needs and requirements of your business application. This involves creating detailed prompts that take into account subtleties within your industry and business to help ensure your model closely matches your goals.
2. List all model options: Evaluate different models based on size, accuracy, latency, and associated risks. This includes understanding the strengths and weaknesses of each model, including tradeoffs between accuracy, latency, and throughput.
3. Evaluate model properties: Assess the appropriateness of model size as needed, considering how model size may affect performance and associated risks. In this step, the focus is on sizing the model to optimally fit the use case, because larger model size is not necessarily better. Smaller models can outperform larger models in your target domain and use case.
4. Test model options: Testing is performed to ensure that the model performs as expected under conditions that mimic real-world scenarios. This involves assessing the quality of the output using academic benchmarks and domain-specific datasets and tuning the model, for example through rapid engineering or model tuning to optimize performance.
5. Refine your choices based on your cost and deployment requirements.: After testing, refine your choice by considering factors such as return on investment, cost-effectiveness, and practicality of deploying the model within existing systems and infrastructure. Adjust your choice based on other benefits, such as low latency or high transparency.
6. Choose the model that offers the most value: Final selection of the AI ​​model that provides the best balance between performance, cost, and associated risks, tailored to the specific requirements of the use case.

Download model evaluation guide

IBM watsonx™ model library

By pursuing a multi-model strategy, the IBM watsonx library provides proprietary, open source and third-party models, as shown in the image.

This provides customers with a variety of choices, allowing them to choose the model that best suits their unique business, geography, and risk appetite.

Additionally, watsonx allows clients to deploy models on the infrastructure of their choice through hybrid, multicloud, and on-premises options, avoiding vendor lock-in and reducing total cost of ownership.

IBM® Granite™: IBM’s enterprise-class foundation model

The characteristics of the basic model can be broadly classified into three categories. Organizations must understand that overemphasizing one characteristic can compromise other characteristics. It is important to balance these characteristics to customize the model to fit your organization’s specific needs.

1. A trustworthy model: A model that is clear, explainable, and harmless.
2. Performance: A level of performance appropriate for the target business domain and use case.
3. Cost-effective: A model that delivers Gen AI with a lower total cost of ownership and reduced risk.

IBM Granite is the flagship series of enterprise-level models developed by IBM Research®. These models focus on trust and reliability and provide the optimal combination of these attributes to help companies succeed in their Gen AI initiatives. Remember, enterprises cannot scale Gen AI with an untrusted foundation model.

View performance benchmarks from our research paper on Granite

IBM watsonx delivers enterprise-grade AI models created through a rigorous improvement process. The process begins with model innovation led by IBM Research that includes open collaboration and training on enterprise-relevant content in line with the IBM AI Code of Ethics to promote data transparency.

IBM Research has developed command orchestration technology that enhances both IBM-developed and select open source models with essential features for enterprise use. Beyond academic benchmarks, the ‘FM_EVAL’ dataset simulates real-world enterprise AI applications. The most powerful models in this pipeline are available in IBM® watsonx.ai™, providing customers with reliable, enterprise-grade AI-based models as shown in the image.

Latest models announced:

• Granite Code Models: A family of models trained in 116 programming languages ​​and ranging in parameter size from 3 billion to 34 billion, in both base models and model transformations on command.
• Granite-7b-lab: Supports general-purpose operations and is aligned using IBM’s large-scale Align of Chatbots (LAB) methodology to integrate new technologies and knowledge.

Try out watsonx’s enterprise-grade foundational model with the new watsonx.ai chat demo. Discover your outlining, content creation, and document processing capabilities through a simple and intuitive chat interface.

Learn more about IBM Watsonx-based models