Begin without machine learning. Use simple heuristics based on content metadata to recommend similar videos to users, and start gathering user event data for future recommender model development.

Start with machine learning. Create video embeddings by training an autoencoder on the content metadata using TensorFlow. Group videos by embedding similarity and recommend from the same cluster.

Launch without machine learning. Present videos to users in alphabetical order, and begin collecting user event data for future model training.

Implement machine learning from the start. Use a public dataset like MovieLens to train a model with Recommendations AI, then apply this model to your data.

Combine both heuristic and machine learning approaches. Use content-based filtering initially and gradually incorporate collaborative filtering as user event data becomes available.