Data is fuel for AI

- In AI, data matters a lot. If you've been tracking AI for a while, or took my previous course, you know the reason for it. AI learns and trains itself with examples, past examples and real-time examples, and those data samples are the fuel that helps AI achieve its objective. The fuel analogy was real. Several years ago, data was even coined as the new oil. The more quality data you had, the better the results you got, the bigger the competitive mode you developed, and the more successful you ultimately became as a company. The latest AI models are challenging that competitive advantage. One of the most impressive aspects of recent generative AI models is that they're already trained on the world's public knowledge. And now that's accessible to everyone, which will make it a commodity over time. Recent developments also enable AI models to create synthetic mock data, or even learn without any data, which is useful in cold start situations, where it's difficult or impossible to obtain examples of training data. These developments lower the barriers to entry for companies that otherwise would not be able to compete. So if you are in this no data camp, that's great news for you. But what does it mean for businesses that heavily rely on proprietary data that is either owned or controlled exclusively by them? Well, if that's you, it's time to rethink your strengths as a business and extend your abilities beyond data capabilities. Data alone will not be sufficient. That said, proprietary data still matters, especially if you can fine-tune the model towards a specific task. Now, you might be asking, "What is fine-tuning," "and why is it important?" AI models like GPT, which stands for generative pretrained transformer, are general purpose AI models, which have a broader objective function and are already trained on all public knowledge. They know as much as possible, and are designed to perform reasonably well in many tasks, just like generalists. ChatGPT, which is the conversational application of GPT, is one example of these models, but you can fine-tune these models to be specialists in a specific area by training them on unique examples you have. And during this fine-tuning process, the model will adjust its parameters to better fit your data while still learning the knowledge it has gained in the past. This allows the model to become a specialist in a specific area that you want it to learn, while still being able to generate in a general way. Imagine taking athletes and training them to be distance runners, or weightlifters, or gymnasts. You can, and probably should do the same with your application. Otherwise, your product will not be very different than anybody else's. You'll be relying on the same capability everybody else is using. Now, you might be thinking that fine-tuning is your answer to maintaining some of your competitive advantage, but it's important to know that creating a private instance of your own fine-tune model can be very expensive and difficult to maintain. We'll talk more about that later in the AI limitation sections of this course.