The original 4chan thread seems to indicate that the leaker verified that his hashes matched with another person who had access to the weights, to make sure that the weights aren't watermarked [0]
Could already have happened in these weights. Reminds me of when the movie studios started projecting random dot patterns during movies to try to catch which theaters were leading to bootlegs. Their approach was essentially defeated by pirates sourcing multiple versions and combining them. In this case, I suspect you could add a small normally distributed random number to some random subset of the weights and it would have very little impact on performance but would corrupt any watermark beyond recognition.
If you find an AI generated response online, and ask GhatGPT if it was the author, it says "it was probably written by a human". But we all know there is a split infinitive here, and an archaic form there, and it knows. But it won't tell us.
