I think it's talking about the weights AIs use. When you feed it all the training data, it gives back weights that are basically weighted probabilities for giving results to prompts. Every word in a prompt gets checked against the weights and overall form a sort of path towards the result you get.

In machine learning, “weights” are the learned parameters (numbers) that inform model behavior. You “train” them by providing the model with examples (data). Each time the model encounters an example during training, the weights update a little bit.

So when Balaji says “train weights” he’s referring to the process of making machine learning models.