(agree but the caveat is "transmission fails -at the tested size- 10-11k examples". given that it works on same distribution, my null hypothesis would be to expect this to work for more deviant distributions with say 6-9 more OoMs of steganized data)

→ View original post on X — @swyx,