I hadn't heard of "Proof of Useful Work". From the the name alone, it sounded like stuff hyc and others were looking at a few years ago, just before it was realized that the only way to really ASIC-proof the algo was a fully random-yet-deterministic state machine (Tevador's RandomX)

But then I did a quick search and read what appears to be the most up-to-date and authoritative paper on the subject, and it's based on the idea that fast matrix multiplication is an inherently good and relevant computation.

Ummm... yeah. No, afaict their paper isn't concerned with the issue that this is an inherently GPU-friendly algo and so favors centralization. The paper focuses of course on making a prover and verifier out of FFM operations.

Which is cool and you can get an ArXiv paper out of it, but not relevant for the real-world at-scale adversary-rich environment that is Internet Currency.

(Edit: 1. No I didn't watch the video; someone please LMK if it discusses a different PoUW 2. I did check and no, Cabanas is not one of the paper authors fwiw.)