We are currently experiencing the AI Summer, witnessing remarkable advancements each day in the field of Transformer models and Generative AI. The introduction of enterprise Generative AI offerings such as OpenAI, Google PaLM, Azure AI, and Amazon Bedrock has created a significant wave in the market. Simultaneously, open-source LLM models like Meta LLaMA, Bloom, Stable Diffusion, and MidJourney etc. have proven to be exceptionally powerful, providing tough competition to enterprise AI companies. Nowadays, launching these models in the cloud has become effortless with platforms like Amazon Sagemaker, Azure ML, and Google Cloud ML.

However, selecting the appropriate AI model for a specific use-case can be a challenging task due to the abundance of LLM models in the market. So let us see how we can build a model selection strategy to choose the right LLM model for a given use-case. There are 3 types of consumers in the AI/ML service landscape –

1. Model Consumers: There are already available fine-tuned GenAI models which are good at some specific industry or use-case like Med-PaLM 2 for healthcare industry or Sec-PaLM for Cyber Security. Model consumers will use these narrow intelligence models as is through APIs or SDKs.
2. Model Tuners: There are some large language models which can do wide variety of tasks like OpenAI GPT 3 or 4. These models excel in various areas, including text generation, text translation, sentiment analysis, and even code creation. Here fine-tuning is required to get the contextual output for our use-cases. Model tuners will fine-tune the model with their own data and will use it through APIs or SDKs.
3. Model Providers: There are some companies who are trying to build their own LLM models from scratch. Model providers will use Cloud machine-learning services like Amazon Sagemaker, Azure AI, or Google Cloud AI to create, train, and deploy their own machine-learning models in the cloud.

The initial step involves determining the category to which our use-case belongs. We can follow the below steps to choose the right model if our use-case falls into the Model Consumer or Model Tuner category.

• If our requirements or use-cases are being satisfied with already available fine-tuned models like Med-PaLM 2 or Sec-PaLM, then we should go for these models as narrow intelligence models which are designed to be good at some specific tasks often outperformed large models which are created to do wide variety of tasks.
• If there are no readymade models available for our use-case or industry, then we should use enterprise LLM models like OpenAI GPT / Google PaLM2 or open-source models like Bloom or RoBERTa and fine-tuned with our own data to achieve contextual response. There are 3 points to consider deciding whether we should go for enterprise models like OpenAI GPT or open-source models like Bloom –

1. Available expertise: It really depends on the expertise available within your organisation to decide whether to go for enterprise GenAI offerings or open-source models as we need ML engineers to provide maintenance and support for open-source models deployed within your organisation network.
2. Data Security: Deploying Open-Source models within your Organisation network is much more secure in terms of data security as your data is not going out of your organisation network. So, we really need to consider the data security aspect very carefully before choosing any Generative AI models.
3. Operational Cost: We need to compare the infrastructure cost for running an open-source models within your organisation network vs. the API usage cost for models like OpenAI GPT3 or Google PaLM2 to decide what is best for your organisation.

Selecting the right AI model for a generative AI use case can be a complex task. However, by considering the factors above, you can start to narrow down your choices and select the model that is best suited for your needs. Model selection is not a one-size-fits-all process. It often involves experimentation and iteration. Be prepared to test multiple models, fine-tune parameters, and explore different architectures to find the best fit for your generative use case. Continuously monitor the performance and adaptability of the selected model as new advancements emerge in the field of generative AI.