the concepts are simple but doing it at scale is hard. there are some good insights on scaling here: * scaling to more engineers/products: IGQL is an interesti

the concepts are simple but doing it at scale is hard. there are some good insights on scaling here: * scaling to more engineers/products: IGQL is an interesting way to compose ML pipelines with straightforward syntax * scaling to more ranking candidates: an active user with a large follow graph who loads the explore tab likely has millions of eligible candidates - how do you load those fast? the idea of using a "distilled" model as a first, light ranking before using a full model as the final predictor is a good intuitive idea that I haven't seen described before. * scaling KNN is hard: FB has done interesting work to make approximate nearest neighbor search fast, and opensourced it (the FAISS library which is referred to in the post). the improvements here are certainly non-trivial. * scaling to more users: creating useful general purpose user embeddings is hard! * scaling to more objectives: instagram has many business objectives, e.g. likes, follows, minimizing hides, so there is a need to have multiple models making many predictions. There is also a need to weight them intelligently, which is where the Bayesian optimization libraries come in. in some sense, nothing is truly AI, but this is useful work which you can learn a lot from.