Hacker News Reader

Rather the other way around: if they don't use a standard algorithm, it's probably beyond useless.

Maybe it's comparable to hiring a dev company that proudly proclaims they use a proprietary programming language that is a trade secret.

So similar to "We use a proprietary data store instead of an off the shelf RDBMS" or "We wrote our own cryptographic library"?

Yes and no. But even for cases, where the algorithm is not published, the general approach can be indicated and hence "Ask them where on Wikipedia you can read more about the approach they took" might be still valid. For ex. you can say you are using a seq2seq model but you don't have to disclose the precise architecture, preprocessing, optimizations and other training nuances.

I was thinking about this one as well. If the process/algorithm/etc. verifiably works on real data for the real use case, even if that's some ripped off the shelf algo or some process combining multiple algos, I think that's fine. Things can be trade secret and not be bad.