PubK are only published if the adress has been used for an outgoing transaction. Of course ~ 80% of the adresses had an outgoing transaction, but for example the big wallets like 1Feex... has never had an outgoing transaction - therefore the pubK is not "published" on the blockchain.
Once again: my goal is not to find wallets with balance. My goal is to find lyrical prefixes.

Now I started something for such an "AI". Yes, it is nonlinear, but with a big dataset and epoches the AI may could crack the whole process, because it is still static, since A -> B. I took 50k pairs of random uncompressed adresses and splitted them into 2 cvs with input and output. I started 100k epoches. If it does not work it is ok, but it is worth a try.