All of a sudden, every company is an AI company! People who, a year ago, would have mistaken the acronym ‘LLM’ to be another flavor of ice-cream claim to be developing foundation models. When you start scratching the surface, their responses would put ChatGPT’s hallucinations to shame.
I don’t blame them. I saw similar behaviors in late 1990s when the world-wide-web had splashed onto the scene and every company claimed to be an Internet company by launching a static website that showed a picture of their office along with ‘About Us’ information. Thirty years later, we still see many trying to complete their digital transformation.
Make no mistake. AI is 10x or more transformational than the Internet was. The challenges everyone is facing is typical of early adoption of any new technology. What are our use cases? What is the ROI? Do we have in-house talent or need to augment? Who are the best service providers? What is the best tech-stack to be used? What happens to our data? Will there be any legal risks? …and many more. This leads to a lot of confusion if not analysis-paralysis. Here are my recommendations, some repeated from my earlier blogs, that should help you start on your AI journey.
Start experimenting NOW! If you have not identified at least handful of use cases to experiment with, hire a consulting organization to help you get started. The intent for early experimentation is to learn and understand what AI can do for you. Don’t fall into the trap of big-consulting companies claiming that they have hundreds or thousands of developers who are experts in AI. You may get better results with smaller or medium sized consulting organizations.
Your data, your models. If you plan to use your confidential data for any AI solution, use open-source models and fine-tune them in your private cloud. Better yet, if you can afford the compute cost then create your custom models. The cost is not as prohibitive as you may think it to be, especially for large companies. This would be the best strategy before we get clarity on T&Cs of data usage and provenance, generated content ownership, copyrights and other finer details from the opaque models from big-tech.
Think ensembles. Whether you call it MoE (Model of Experts), Model Clusters or Model Ensembles, think about fine-tuning a set of open-source models, each focused on a specific task or topic with an orchestration model to manage user interface. This will give you flexibility to try various models based on use cases and replace a model with something better on need basis.
Governance is imperative. AI Ethics and Data Governance should not be an after-thought. This should be established up-front with clear metrics on model evaluation, data retention, provenance & lineage along with many other decisions your team will need to make as part of data pipelining.
I have touched upon some key recommendations that should help you get started. Although this article is focused on GenAI vendor selection, it highlights some key points that you should take into account for your own AI transformation strategy.