EDIT: Maybe I'm just thrown off by the "Powered by AI" part of the article title. I was expecting more I suppose.
There’s really no point in having a semantic argument, but just know that if you wish to do so you are arguing against many decades of wide usage of the term.
* 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.
What makes this mildly interesting is the IGQL, but again, without knowing the full syntax, it feels pretty restrictive.
Youtube is doing much more advanced stuff, like Reinforcement learning@Scale, as comparing to Instagram in this regards.
"look Ma i'm writing my own AI algorithm!!!" ... "writes linear regression by hand in python"
A. System should, without prompting, identify areas of improvement and innovation
B. Automated collection of data and the processing thereof, combined with application towards a concrete goal — does not qualify under A.
C. Part of A. is willing and unwilling discovery and exposure to both benevolent and adversarial environments and operating conditions
Bengio has a paper in 2003 that describes almost the same idea as word2vec (CBOW model to be exact).