Have you ever wondered what makes a successful GenAI project? Is it the model, the data, the tech stack, or something else? We have been working on various GenAI projects already and along this journey, we have discovered that the most important factor for GenAI’s success is often overlooked – the team crafting solutions. They are the ones dealing with new tech, sorting through the noise, and finding the right path for their organizations. But they also face many challenges that are not technical, but human. These challenges can affect the motivation, creativity, and productivity of the AI team. In this article, we will reveal what these challenges are and how to avoid them. We call them the “Five Anti-Patterns in GenAI Research and Development that Set You Up for Failure.”
Anti-Pattern 1: Getting Paralysed by Overthinking and Over-Defining
Despite the hype around Gen AI, putting them into production remains tricky. Even if you as the AI team have the support of the company’s board, increasing their confidence in the investment is crucial. So dive boldly into the first Proof of Concept (POC). Begin coding as soon as you have confidence in an idea.
Don’t wait until you have all the details and answers. Gen AI is changing fast, and you might miss the chance to innovate. Be ready to face the unknown and take some risks. Accept the messy and unpredictable nature of the initial POC as a natural part of the process.
The goal is not to know everything, but to build confidence by doing. For sure, your first POC will not be flawless, but it is the crucial step needed to build a foundation for future success. Be flexible and adaptable in your AI development and avoid the anti-pattern of waiting until everything is perfectly defined before taking action.
Anti-Pattern 2: Assuming Your Awesome POC Speaks for Itself
You have completed the first Proof of Concept (POC) for your Gen AI solution. Congratulations! But don’t stop there. After completing the POC, it’s not enough to sit back and expect others to recognize its value, It is on you as AI lead to to move beyond the technicalities and communicate the value of your POC to others, especially the decision-makers in your company. They are not as familiar with the technical details as you are, and they are bombarded with AI hype every day. How can you make your POC stand out and relatable?
The answer is value-focused storytelling. You need to craft a compelling narrative that showcases the strengths, utility, and real-world value of your AI innovation. You need to explain what problem it solves, how it solves it, and why it matters. You need to use clear, simple, and engaging language.
While it may be tempting to dive into the complexity of the process to show how difficult it was and how much you suffered, resist that urge! Remember, your role is not just to showcase the AI’s capabilities, but to make its use case understandable and compelling to a broader audience within the company, so keep the daunting details to yourself. Clean up any internal chaos before presenting the solution externally. Make the complexities appear simple, and only go into technicalities when specifically asked. Remember, your role is not just to showcase the AI’s capabilities but to make its use case understandable and compelling to a broader audience within the company.
Anti-Pattern 3: Falling Into the Hasty POC Sequel Syndrome
You have gained the confidence of your stakeholders with your first Proof of Concept (POC) for your Gen AI solution. Well done! After all, output is what matters, right? So, wanting to generate another POC means desiring another output. But don’t rush into the next POC right away. You need to take a strategic step back and spend some time to reflect on what you have learned and how you can improve.
Your team has worked hard and faced many challenges in the first POC and you all have collectively learned from the mistakes made in the initial phase. You need to acknowledge their efforts and feedback. You need to conduct a detailed post-mortem meeting involving all teams, It is crucial to conduct a detailed post-mortem meeting involving all team members, not just to release some steam, but to structure and agree on the next POC delivery process.
You need to gather lessons, define processes, and share knowledge. You need to streamline the workflow from the discovery of the POC idea to development, evaluation, and stakeholder delivery. You need to make everything transparent and well-defined, especially the parameters of the POC and the scope of deliverables.
Remember, the first POC was a learning experience. Don’t make the mistake of jumping into the next challenge without thinking about the journey so far. Clean up any internal mess, and make sure the accumulated knowledge becomes the foundation for a more organized and efficient process. Only after this thorough assessment should you embark on the next phase of experimentation and development.
Anti-Pattern 4: Neglecting Team Harmony by Ignoring Role Dynamics
Teamwork makes the dream work, especially in Gen AI. You need to have the right people in the right roles to make your Gen AI solution a success. But who are the right people, and what are the right roles?
There are three main categories of roles in a Gen AI team: data scientists, developers, and the product team. Each category has different sub-roles that require different skills and expertise.
The term “data scientist” often oversimplifies a diverse skill set. Depending on the company’s goals and the specific Proof of Concept (POC) at hand, different flavors of data scientists may be needed.
For example, suppose you want to use pre-trained language models (LLM) for your Gen AI solution. You can either use existing APIs that provide access to LLM, or you can train or fine-tune your own LLM from scratch. Depending on your choice, you will need different types of data scientists. If you use APIs, you will need data scientists, who have software engineering skills and can integrate LLM into your solution often refers as ML engineers. If you train or fine-tune LLM, you will need ML researchers, who have deep learning and mathematics skills and can create and optimize LLM for your specific problem.
You also need to consider the roles that are changing or emerging in the Gen AI era. For example, quality assurance testers are not just checking for bugs but also evaluating the user experience and value of the Gen AI solution. Product managers are not just defining the requirements, but also understanding the Gen AI capabilities and more importantly its imitations. Prompt engineers are a new role that is responsible for designing and optimizing the prompts that interact with the Gen AI models. You need to define the responsibilities and skills of these roles and make sure they are aligned with the rest of the team.
Having the right roles is not enough, though. You also need to have the right team harmony. You need to create a culture of collaboration, where each role respects and supports the others. You need to foster a spirit of synergy, where each role adds value and enhances the others. You need to promote a habit of communication, where each role shares insights and feedback with the others. You need to cultivate a space of harmony, where each role works together to create a better Gen AI solution.
Anti-Pattern 5: The Mirage Trap – Unmasking Unrealistic LLM Expectations
Large language models (LLM) are powerful tools that can generate text or images. Some examples of text-to-text and text-to-image LLMs are GPT and DALL-E models by OpenAI. These models are widely available and easy to use, which makes them appealing for everyone to experiment with. Many companies organize hackathons to encourage their employees to create something with LLM and solve a problem quickly. These fast prototypes can be very inspiring and fun, but they can also create a false impression of what LLM can do in production. Building a prototype that works for a weekend hackathon is not the same as building a product that works reliably, safely, and ethically in the real world. LLM are not magic; they have limitations and challenges that need to be addressed. This is why many companies have not released any actual products based on LLM, despite showing impressive demos. Therefore, it is important to manage the expectations, educate the people iteratively about LLM, their strengths and weaknesses, and accept the fact that not all problems can be solved with LLM, nor should they be.
Last thought:
GenAI research and development is an exciting but challenging journey. You need to be bold and curious, but also strategic and careful. You need to avoid the common mistakes that can derail your GenAI projects, such as overthinking, assuming your POC speaks for itself, rushing to the next POC without reflecting on the lesson learned, ignoring the team skill set and harmony, and falling for the mirage trap.
As we enter the new era of AI, we need to create a culture that supports GenAI innovation. We need to be adaptable, transparent, and eager to learn. We need to build GenAI teams that are strong, diverse, and collaborative, and most importantly we need a team to turn our GenAI ideas into real solutions that make a positive difference in the world.
