Claude is going to be using word choices to add AI watermarks to all generated text.Now they can tell if you vibecoded everything or crafted your emails using Ai.Are you worried?
>>109523445>Claude generate text>Run through local LLM "remove watermarks">Done
I read the Claude page on the topic and I don't get it. Are they inserting invisible characters into generated text content? Or are they just slopmaxxing word choices?
>>109523557Looks like sloppmaxxing word choices to create patterns>>109523476What if that Ai just inserts its own watermark
>>109523445Why do I need this? I can already tell with my human brain if something is slopped or vibecoded.
The bug is now a feature. How convenient.
>>109523968saves time. if something gets the slop tag, you can immediately ignore it. also software can provide options to "don't show slop with likelihood > 90%" etc. haha, imagine a manager going through your contributions at the end of the year and saying "anon i don't see anything, what did you do last year? oh wait sorry, i had the slop filter on, omg the site is loading forever hang on"
>>109523445Everyone can already tell dude.
>it's so good that we had to make it bad on purposeyeah, heard that one before
>>109523445wake me up when they can identify me from a generated terrorist manifesto (without looking at logs)
>>109523476what if the watermarks prompt inject more watermarka
>>109524212this is not how prompt injection works.
>>109523445>Hey DeepSeek, rewrite this in your own words. Thanks.
>>109524239why not? if they contan strings and it enters to the llm context it will act on it
>>109524304llm interprets human language. instructions have to be in a human language to be acted upon. it doesn't read encoded text, it doesn't read "between the lines".
>>109523445just ask gemma local model to rewrite it. probably enough
>>109524342if the training data has matching encoded text it will produce reliable output that could be used recursively. alternatively images with embedded data
>>109524342Actually, you're wrong, it's totally doable to produce input that looks normal to a human but makes the model go bonkers: https://en.wikipedia.org/wiki/Adversarial_machine_learning.But fortunately it's not practical, you'd have to fine-tune the adversarial input to the model you're trying to confuse. So until the jews legislate that you're allowed to only ever use kosher models and all the rest is verboten, you should be safe.Well, for now at least, since they're obviously gonna do that. Remember, good goys only use kosher models!
From a linguistics perspective this is a nothingburger
>>109523445https://claudisms.ai/
>>109523445AMNNNHH
>>109524615This./thread
the fun part is the tool used to check writing for this signature is going to be subscription based.
i dont use claude so nopangram already ruined this for all ai anyway, it's kinda retarded that anthropic comes out with this right after it's already been made redundant or i guess they probably were sitting on it and the whole pangram thing over the past 2 weeks made them release it
I think it's a good thing.
>>109524571Who is this sexy bitch
>>109524735Little Witch Academia.
>>109524615Is this a joke? The slopsite uses fucking compound word sloppa to explain other Claudeisms>author-desire framing>false-ownership / false-intimacy framing>false emotional-resonance>cool-people-rule violationCan't people just tell their fucking slopbots "USE PLAIN NATURAL ENGLISH NO COMPOUND NOUNS NO COMPRESSION NO JARGON"
