MLGonzo
Region: GB
Monday 17 August 2026 18:46:39 GMT
7571
333
23
10
Music
Download
Comments
Caleb Solano :
But this method would generally result in watermarks that only the model that produced it would be able to recognize, right? So only Claude would be able to tell you if Claude wrote it?
2026-08-18 03:18:46
3
Braaamskii :
I understand text, what about code, are they watermarking the code too?
2026-08-17 19:43:45
6
kue lapis legit :
worth knowing. been using walter writes since the watermark news, it rewrites past it and the meaning holds
2026-09-09 07:56:42
0
Neil Owen :
This only works if you can produce the original prompt right, because that’s how you get the probability distribution? That makes it pretty much useless.
2026-08-18 13:10:19
3
Matthew DeKrey :
Tokens' letters are mapped from numbers, right? I would use a filter on the possible tokens every handful of output tokens, such that maybe every 5th token needs to have a number that is modulo 4 = 0. This would prevent the need for rebuilding the ranking, but would be straightforward to verify if it came from a particular model afterwards rather quickly.
2026-08-18 11:52:20
1
TG :
Okay… but how do we do this in cobol magic code man?
2026-08-18 02:17:04
2
Chris Bowles :
brb, running to post Cat on LinkedIn
2026-08-18 13:36:22
1
Luke Owen :
Does this not presume that all sequences of the next token are the same? Surely there’s some variability to that? Or will there be some accepted %
2026-08-17 22:38:08
0
stephen.purkiss :
From the little I did read about it I thought it was simpler in that you know how you can have a seed for a random number generator because computers can’t actually generate random numbers, so they manipulate the seed that figures out the response and the result is you know what the response is going to be because it slightly changes the words that are returned and that’s where people are having an issue because you’re essentially tampering with the “actual” response the AI would provide without any manipulation.
2026-08-17 22:34:25
1
tony :
Never mind the watermark, where can I get a "The cat sat on the aubergine" T-shirt ??
2026-08-23 09:40:56
0
John Hess :
maybe
2026-08-18 22:39:06
0
cah :
My concern is they're degrading the code output to something less than optimal.
2026-08-18 01:49:08
2
David Jackson :
How are the numbers hidden in the text. If you are just copying the Unicode characters
2026-08-17 20:44:12
1
Thomas Watkins :
if we can prove ai via encoding a hash into the output, can't we just have the input and then pick those tokens? completely bypassing the need for weights. I agree that intentionally choosing the next token is a great method of encoding "42" every 42nd token. but it's got to be far more deterministic and more resilient than a simple hash.
2026-08-17 20:45:47
0
Harper :
In formal logic parlance, the absence of a watermark makes the material implication vacuously true.
2026-08-18 03:03:42
0
Imiix Solar | Digital Artist :
Can happens with imagen this?
2026-08-18 15:49:32
0
AI Fan Token — $AI :
Transform
2026-08-18 04:23:01
0
To see more videos from user @mlgonzo1, please go to the Tikwm
homepage.
