The One Practice That Is Separating The AI Successes From The Failures

The One Practice That Is Separating The AI Successes From The Failures

Somewhere between 60-80% of AI projects are failing according to different news sources, analysts, experts, and pundits. However, hidden among all that doom and gloom are the organizations who are succeeding. What are those 20%+ of organizations doing that are setting themselves apart from the failures, leading their projects to success?

One of the biggest insights from these AI successes is that they don't see AI projects as application development or functionality-driven projects. Rather, they see them as data projects, or sometimes even data products. A data project doesn't start with an idea of what the functionality needs to be, but rather focuses on what insights or actions need to be gleaned from the data in whatever current shape it's in.

Agile, the most popular methodology for application development, falls flat when dealing with AI because it doesn't tell you how to deal with data, the core asset of an AI system. That's where the CPMAI methodology comes in. It updates the CRISP-DM model with Agile and AI-specific details, providing a framework for successful AI project management.

Retrieval augmented generation (RAG) is an exciting new use of large language models (LLMs) that could provide the benefits of fine-tuning without the need for retraining. RAG allows the LLM to produce more specific responses by incorporating additional information from the context. By tokenizing the question and using these tokens to find snippets of documents, RAG can create prompts that enhance the LLM's ability to generate accurate answers.

Fine-tuning an existing LLM is another option for improving its performance. By training the model further with additional information, it can be updated without starting from scratch. This approach is particularly useful when there is a need to incorporate domain-specific knowledge or when training on sensitive data. However, it's important to note that fine-tuning can become expensive if done frequently.

If you're looking for a quick way to add generative AI capabilities to your application, using the APIs from a public LLM is a viable option. Major public LLMs offer APIs that allow you to access their models without the need for hosting or additional infrastructure. However, it's essential to understand the expected usage and estimate costs before pursuing this option.

Training a large language model from scratch gives you complete control over the information the model sees. This option is ideal when you have a specific domain or industry knowledge that can be leveraged to create a domain-specific LLM. It's also suitable for organizations that are heavily regulated and require full control over the data used.

Investing in all phases of the AI/ML workflow is crucial for success. From storage solutions that can handle large datasets to tools for preprocessing, feature engineering, model training, testing, and model serving, using the right tools for each phase will speed up the process and produce better results. It's important to build data pipelines and model-training pipelines efficiently and track results to improve models' performance.

When starting with AI/ML, it's recommended to begin with simple problems and low-hanging fruit. Building models that make simple predictions, such as regression, categorization, or classification, requires less compute and investment compared to large language models. This approach allows organizations to see value and build the necessary infrastructure and processes for more complex projects.

In conclusion, successfully separating AI successes from failures requires a shift in mindset towards viewing AI projects as data projects. By adopting methodologies like CPMAI and exploring techniques like retrieval augmented generation, fine-tuning, and leveraging public LLM APIs, organizations can increase their chances of AI project success. Additionally, investing in the right tools for each phase of the AI/ML workflow and starting with simple problems will set the foundation for future AI endeavors.

Actionable advice:

1. Change your perspective: Instead of treating AI projects like traditional application development, approach them as data projects or data products. Focus on extracting insights and actions from the data.
2. Embrace AI-specific methodologies: Agile may work for application development, but it falls short when it comes to AI. Explore methodologies like CPMAI that incorporate Agile and AI-specific details to effectively manage AI projects.
3. Start small and build on success: Begin with simple problems and low-hanging fruit to gain experience and demonstrate value. This allows you to develop the necessary infrastructure and processes for more complex AI projects.