[vessenes]

FWIW, Phil Wang (lucidrains) has been working on Titans reimplementation since roughly the day the paper was released. It looks to me from the repository that some of the paper's claims have not been reproduced yet, and reading between the lines, it might be Wang considers the paper to not be that groundbreaking after all -- hard to say definitively but the commit speed has definitely slowed down, and the last comments involve failing to replicate some of the key claims.

Unfortunately. The paper looks really good, and I'd like for it to be true.

https://github.com/lucidrains/titans-pytorch

[codelion]

It's a shame when potentially interesting papers don't hold up in practice. I've seen a few cases where the real-world performance didn't match the initial claims.

[thunderbird120]

Yeah, that's the normal outcome for papers like this. Papers which claim to be groundbreaking improvements on Transformers universally aren't. Same story roughly once a month for the past 5 years.