Maybe first try on a small subset of the data just to see if it actually works? I.e. avoid mem restrictrion

Is this a sort of logic? LIke 0 1 0 1 0 -> next is 1.

How would you expect that any of these 100k binaries has a predictive power for the next one? If there are sequential patterns predicting the next value, try an LSTM

I don’t see where c++ and java were applied.

Also I would ask why you would try to build things such as kNN from scratch. Evidently does not add anything over existing implementations.