Gen AI Model Selection and Development

We've learned that large language models or LNS are especially effective in handling and analyzing qualitative, unstructured data within project management workflows.

We also learned that the cost of developing an LLM from scratch is very high, therefore making this option unavailable to most organizations. If your organization needs to sel ect and tailor a gen AI model, your options will likely look like this:

Use a public general purpose model like the chat GPT you are familiar with, or GPTs, purchase a purpose specific SaaS enterprise AI solution. Use or customize a cloud-based secure LLM. Let's quickly walk through the pros and cons of each.

Public general purpose model Pros & Cons

Pros, cost-effective, and often involves lower or no initial costs. Easy, quick deployment. Since the infrastructure and model are already set up, regular updates and maintenance are handled by the service provider. Scalable according to the needs of the business without worrying about the underlying infrastructure.

Cons, less customization as the model is designed for general use, potential privacy and security concerns as data is processed through a third party service. Dependence on the service provider for uptime and performance.

A purpose specific SaaS enterprise solution Pros & Cons .

Tailored to the needs of larger organizations with more robust features. Typically includes dedicated support, may offer better integration. Capabilities with existing enterprise systems often comes with enhanced security and compliance features suitable for enterprises. Cons, higher cost due to premium features and services may include features that are unnecessary for the organization. Adding to complexity, your organization might become locked into a particular vendor. Use or customize a cloud-based secure LLM Pros high level of security. Ensuring proprietary data is kept confidential, more control over the model and its usage, including data handling can be customized and trained on specific data sets to better serve the organization's needs. Cons, typically the most expensive option due to licensing fees and the potential need for dedicated infrastructure requires more management and oversight from the organization's IT team may involve longer setup and integration time.

AI infrastructure costs are high due to the complex computational demands of tasks like generating text with GPT-4, which far exceeds simpler operations like sorting databases.

Transformers allow us to predict their compute and memory needs, which helps in the selection of suitable hardware. Despite advancements in GPU speeds and training optimization, the growing demands of AI development continue to drive up infrastructure costs. To minimize expenses, select the least complex model that meets your needs. For detailed training, cost calculations, refer to the navigating the high cost of AI compute article.

At the end, when choosing and tailoring an LLM align your decisions with your project goals and available resources to leverage LLMs effectively for your specific applications.