We keep hearing about how AI systems are trained on large volumes of data…but how exactly are AI models trained?
To oversimplify the process, there are two main ways AI systems can learn:
- Supervised
- Unsupervised
Supervised learning uses labeled data with known outputs. For example, to train an image recognition model, you might feed the AI system thousands of images labeled with the correct categories (e.g., “dog,” “cat”). The AI learns to map inputs (the images) to outputs (the labels) by identifying patterns in the data.
Unsupervised learning uses unlabeled data. The AI has no direction and must find patterns or structures on its own. A common method is clustering, where the model finds features about the data and groups items with similar features together. Keep in mind that websites are an unlabeled dataset.
There are other ways AI systems learn. Reinforcement learning involves training an AI to make decisions by rewarding it for successful actions and penalizing failures. There is also self-supervised learning where the AI system generates labels from unstructured data. We also have semi-supervised learning where the system uses a mixture of supervised and unsupervised learning methods. It gets complicated fast, obviously, but if you keep in mind supervised vs. unsupervised learning you’ll have a pretty good sense of what’s happening.
To take ChatGPT as an example, they start with unsupervised learning and then fine-tune the model using supervised learning (though the supervised aspects are limited and on small datasets from what I’ve read). ChatGPT also uses reinforcement learning from human feedback (RLHF if have ever seen that acronym used) to help further refine the model. I think it would also be safe to assume the learning process is more involved than anything we’ve been told publicly.
With all this in mind, we can start to understand how to optimize our websites for generative AI search. We need to improve clarity, completeness and accuracy to help the model learn more effectively with unsupervised training. For example, providing more examples gives the AI system more context so they can group our page correctly. As another example, explaining concepts in different ways can help the AI system learn different patterns. Having more depth will make our website a better resource to draw from than a website that only offers vague information.