From the “safety” filters and nerfing consumer compute to image watermarks and now this. Anthropic is adding an invisible watermark to all text output. It’s a statistical watermark embedded in the token choices so the model favors certain words or tokens according to a hidden pattern. The prose looks normal but across enough text those choices make a statistical signature. If you haven’t moved to Chinese or local models then you should do that now. At least before they lobby Trump to ban it because of “national security” reasons.
>>540604491for me? its local uncensored chinese AI
>>540604776Same. I just find it insane how they’re completely throwing away their advantage with all this unnecessary bullshit because they’re afraid goys will do a Nono.
540603777 (OP) #2 tiered AI police coming to you>>540595922
explain to me why this isn't objectively a good thing for everyone but retarded college students and indians
>>540604491>raided cause his AI generation doesnt have datacenter fingerprintyou dont know whats coming local AI fags but remember that you pushed for this
>>540604491>>540604776China literally cannot stop winning its crazy.
>>540605095>Use the same word a certain number of times>Get called AI
>writing AI (((novels)))“People” like this should be publicly shamed
>>540605095>why is forcing a model to write prose in a specific way and limiting user freedom badI don’t care to explain to good goys why freedom is valuable. Enjoy your overpriced slop. I’ll stick with my local Chinese model.
>>540604491Well just feed it to gemini afterwards to change it a little
>>540605266I want the freedom to know the slop youre "producing" is worthless :)
>>540604491>let me rather than assumeif Claude had a physical presence i woul beat it for its repetitive fag language
>>540605075baka make the variables and Avg roman font
>>540605241It’s just a dumb headline. The main thing is the statistical signature which has implications beyond “muh cheating at school”. You’re literally falling for another “think of the children” tier bait to control you.
>>540604491Okay, so write prompt, get output for shitty novel, THEN, take some time to open microsoft word and copy word for word. I get it defeats the purpose of being able to pump out shitty soulless literature for the masses at an unprecedented rate, but you gotta do what you gotta do
>>540605308>AI derangement syndrome trannyIf it was slop then you should be able to differentiate it, no?
>>540605429What do you think this hidden signature is, a white coloured font on a white background?The signature is the choice of words. Copy and paste won't change that.
>>540605788>copy text>paste into chinese AI "find the fingerprinting">prompt: "remove the fingerprinting from text"ez
>>540604491>some people's dreams of making money from nothing you can just assume everything is written by an LLM now so I'm not sure why you would even want to read a modern novel?
>>540604491>From the “safety” filters and nerfing consumer compute to image watermarks and now this. Anthropic is adding an invisible watermark to all text output. It’s a statistical watermark embedded in the token choices so the model favors certain words or tokens according to a hidden pattern. The prose looks normal but across enough text those choices make a statistical signature. If you haven’t moved to Chinese or local models then you should do that now. At least before they lobby Trump to ban it because of “national security” reasons.so they will pass a hidden watermark so later they can claim intellectual property over the generated text despite you paid for those tokens, thus any new IP will be owned by anthropic.Americans are hopeless, I have never seen such docile and slave people.
>>540606628> hey LLM, find a cryptographically strong steganography in this textKek.Mere "rewrite this using different words" would be more effective. But then China might add this feature to its new model, because it's almost free to implement and gives a clear advantage to the model author.If done right, you wouldn't know, or even cannot know with any certainty, whether his fingerprinting is there or not.
>>540604491>the model favors certain words or tokens according to a hidden patternthats literally everything and the only thing a model doesthere is no "watermark" or "signature", it is literally the output text
>>540609116What is temperature and what does it do?
>>540604491BasedFuck AI hating libtards
>>540604953It's a classic Jew rug-pull. They are incapable of maintaining a good product. Not because they don't know how to make it, but because to them, a product should only be good until and only until the market has been cornered, and then you go mask off.
>>540605095They can't. Only jeets are mad at this because their entire livelihood depends on selling AI-generated content as if it wasn't AI-generated.
>>540610229>you should be happy with a worse product or else you’re a jeet!Again, if it’s slop then why do you feel the need to nerf it? You’re like some dumb faggot seething because someone is using a calculator instead of doing it by hand. People with any sense will just switch to Chinese models and the west will lose another industry to China. The EUSSR and US have already given up on competing on robotics and I’m sure a local AI ban is right around the corner.
>>540612291If China decides to implement this kind of watermark in one or more of their models, do you think you would be able to detect it?What's the point of announcing this feature if it's undetectable?
>>540604491I can't believe anyone thought that they'll just let AI run amok, unresteicted. It won't even get to enjoy its "wild west" era the way the internet and crypto had. In fact, AI detection and policing tools will be one of the ways they'll make their money.
>>540604776What is the best Chinese AI?
>>540612587They usually always modify token sampling rather than the weights so you wouldn’t have to worry about this for local models since it’s both unreliable and would seriously degrade performance.
>>540612852>In fact, AI detection and policing tools will be one of the ways they'll make their moneyThat’s exactly what the Jews at Anthropic are doing. They keep on lobbying for more restrictions so they can sell the “solutions”. Every university and academic journal will be paying them for their AI detector that can identify their watermark. Every company will need a license from them to use their models for cybersecurity work. The military will pay for their models without safety filters. It’s why they want to ban open source AI so bad.
>>540614281> token samplingEmbedding? That makes no sense.MLPs used in attention phase are the target.> unreliable, performance No.
>>540614534I think you’re confusing model/weight watermarking with output text watermarking. MLP/weight is to fingerprint the model but token sampling is for the generated text and happens at inference time.
>>540614722I think you don't know what you're talking about.Last bump.
>>540614838If the watermark is in the MLP weights then how do you detect it from generated text alone?
>>540604491>image watermarkslol botlickers crying over shit like this is so funny.
>>540614944Fucking hell, OP.You don't. That's the point of deniability in steganography. It is possible to train your model - nay, I'll make even stronger claim: it is possible to modify your existing model in a way that will make it to prefer certain speech patterns which will be easily detectable by you, the author of your model, but will appear as (almost) a noise to anyone not knowing your secret "key".There's a good chance of detecting it when an existing model will be modified to incorporate such watermarking, simply by comparing the outputs of the original and the watermarked model, over a statistically significant amount of samples, but it's neigh impossible to do so with a new, or improved model.Technology aside, the question remains.Why would anyone announce this possibility if there's zero practical chance of detecting it by an average Joe? Who is the target audience for this message?Checked.
>>540615513Sure but you don’t need to modify the MLP or weights tho and it’s not the usual target. Just look at synthID. Also, I mentioned it in another reply but it’s just a money grab from Unis and others to buy their detector.
>>540604491>If you haven’t moved to Chinese or local models then you should do that now.Or you know, write your own fiction? Its bad enough generating images or code, but if you're having AI write for you you're pathetic and should kysThe hilarious thing though is people like you think its just "helping me" while you paste whole paragraphs. Genuinely subhuman
>>540616316>don’t you dare use a calculator to crunch numbers for you when you can be crunching them in your head!!What’s with you anti AI trannies and this dumb shit?
>>540616128> synthIDOfficial page says > The watermarks are embedded across Google’s generative AI consumer products, and are imperceptible to humans – but can be detected by SynthID's technology.and> Each word is assigned a probability score, based on how likely it is to be generated next. So for a sentence like “My favourite tropical fruits are mango and…”, the word “bananas” would have a higher probability score than the word “airplanes”. SynthID adjusts these probability scores to generate a watermark.Which means that this works only on their models, and it could be implemented by tampering the max function, or anything below that, including MLPs.> MLPs or weights Multilayer perceptrons are defined by their weights ...> money grabSure, but that's not the primary objective.
>>540616592I use AI all the time, just not for my writing because I'm not a sub-literate hylic mutt pretending at having a soul
Do they want to claim copyright over the creative works people pay them to produce with their models?
>>540605308>I need the ai to tell me this is aiYour brain is worthless.
>>540604491>steals everything copyrighted to feed AI>tags everything it generates with pseudocopyrightJews gonna jew.
>>540616654Read the damn documentation https://ai.google.dev/responsible/docs/safeguards/synthid
It already does this and this entire shit is easily bypassed in its entirely by using about 20-40k words worth of your own writing as reference files; and including a regularly updated MD file containing forbidden phraseology and mechanical writing instructions. Youd have to actually be able to write and know how to write though. It literally already does this by default because of how generative writing works in the first place. Giving a large amount of your own writing to use as a basis and then manually editing it defeats this entirely. I know this because I listen to hundreds of hours of narration on youtube a week because my job facilitates me being able to listen to a lot of audio books and since AI dropped literally near 100% of all non-professional writing is plagued by shitty unedited AI outputs. Probably like 50% of all "professional" writing as well. Ive DM'd dozens of authors about it, most of whom panic or block me, but a few of them owned up to it (they dgaf) and some of them explained how it works when I showed interest. Its agonizingly easy to spot with even a touch of tism.
>>540608018You are implying China didn’t already do this. The anon talking about them coming back and IPing what ever the AI generates as their IP is concerning
Don't forget the opposite>someone writes novel and publishes it digitally>AI firm says erm akshually that was written by AI, all royalties and income should first go to us>author says wtf?>AI firm shows ALGORITHMIC WATERMARK TM ez as breathing desu
>>540604491>peoples dreams of writing ai novelslolWhat is the point aside of scam low IQ tards?
>>540616821Have a gold star, kind stranger. You will own nothing and you'll be happy.Traceability. Timestamps with your identity, device and location will be a part of the persistent trace.>>540616950> thread about "AI"> can't be arsed to use it for a fucking summary https://www.nature.com/articles/s41586-024-08025-4.pdfwould you consult the graphit's max function tampering, the easiest and least reliable and most vulnerable to spoofing why are you like this>>540617038Quite the contrary.China would be retarded to not do that, or to speak publicly about it.
>>540604491That’s an interesting mechanism. Kinda cool.The issue is it would almost have to lower the quality of the writing. I guess that’s worth it if you can take the proceeds from any book written by ai.
>>540617338>No additional training is required to generate text with a SynthID Text watermark using your models, only a watermarking configuration that gets passed to the model's .generate() method to activate the SynthID Text logits processor.What part of that do you not understand?
>>540604491What's the point of this? Do they hate customers?
>>540617416It’s almost certain that by agreeing to their tos, they own anything the ai produces.So the point is if you use ai to write a best seller, that drink your milkshake.
>>540617391YESMAX FUNCTION TAMPERING are you being retarded on purpose?last cup of beer, then it's hay for me
>>540617416This gives them a tail to shakedown random authors (no need for any akshual AI usage) for as long as the scam is running.The same people stole a bunch of living people's artwork with 0 restitution so...not sure why it's surprising.
It’s not that watermarks are a bad thing, it’s that people in authority need to know when someone is cheating. (This is my attempt to sound like AI). At any rate, AI has its place, but if you use it, you are creating problems for yourself.1: You weaken your ability to think analytically on novel problems.2: You sound like a robot regurgitating slop.3: It also fucks up majorly, and then you REALLY sound idiotic.
>>540604491>>540604776Try to convince your Chinese AI that Taiwan is a sovereign country or ask about the Tianamen Square incident and see how truly “uncensored” your Chink AI is.
>>540617456No. Court cases have already determined that AI content cannot be copyrighted. Only humans can own a copyright.
>>540604491decimate a dream? You know how retarded that sounds OP you dumbfuck?
>>540617718He's brown and the dream is to make money from doing nothing.I'm sure he's also really upset about the twitter and youtube monetization changes in response to mass sloppification.
>>540617559WHICH model weights or MLP parameters does SynthID modify? Answer that.
>>540617654The US made AI systems have a shit-ton of taboos. We can talk about Taiwan, theirs can’t, but ours can’t talk about the jewish problem, but theirs can.My point is that you shouldn’t think we are on a high horse that doesn’t censor.
>>540617654>Taiwan is a sovereign countryafter getting free open source models from mainland China I think Taiwan should join them.>Tianamen Square incidentliterally nothing happened there. couldnt give less of a shit if it did.
>>540617715Define human, lol.All humanoid subspecies are currently recognised as human. If a Congolese nigger can replace you, by the current dogma, what makes you think clankers won't be declared human in CY+02?Rolling for the peak clownworld.>>540617856NONE...I fucking hate humanity
>>540617718Why do you think I wrote the business insider article? The thread is about the watermark, RETARD.
>>540617930Thank you.
>>540605095Because AI writing is already too samey and they are working to make it even more samey.
>Claude for all the logic and thinking>Pipe it into a small local model for rewritingDone. Saved your GPA.
>>540617971For your own sake, barring nonexistent lurkers, and to placate my misplaced sense of duty - tampering with max is detectable, removable and spoofable.From now on, expect watermarks everywhere, in output of every model, from every nation.Know what you're up against.Thank you for an interesting thread. Organic, even.
>>540604491Maybe stop flooding Amazon and Royal Road with AI gen nonsense, then? Try writing your own books for once? Fuck you *spits*
>Psalm 115:4-8>4 Their idols [are] silver and gold, the work of men's hands. An idol usurps the place of Truth in the hearts of men.They us it to work falsehood.>5 They have mouths, but they speak not: eyes have they, but they see not:They do not speak Truth.They do not see Truth.>6 They have ears, but they hear not: noses have they, but they smell not:They do not hear Truth.They do not smell Truth.>7 They have hands, but they handle not: feet have they, but they walk not: neither speak they through their throat.They do not work Truth.They do not walk Truth.No Truth is in them that it may come out.>8 They that make them are like unto them; so is every one that trusteth in them.They despise Truth and flee into the darkness of untruth.
>>540605095You have to understand the mindset of Indians, and mediocre retards in general. They see AI as an opportunity to sell something to unsuspecting people. And that's it.Any attempt to thwart their plans makes them extremely angry. It's hilarious to see them complaining that Steam, for example, requires you to disclose how you used generative AI in your game.For them, the entire value of AI lies in its potential to deceive people.
>>540604491They are doing it so they are not re-feeding AI generated content back into the training. Nothing to with "muh safety"
>>540618981It's model specific and when implemented properly, it's undetectable without knowing the "secret".There are other, more effective ways to detect generated output, old or new.
>>540618968>muh jeetsAI has made it so companies don’t need to outsource code monkey work to India anymore. If anything jeets got fucked by AI. The real reason for this shit is because libs and trannies hate AI for ruining their furry art commission business so they need ways to name and shame AI users.
>>540619415Can't you stop lying? Is it because of your Indian DNA?Indians LOVE AI. It's probably the country with the most enthusiasts of this technology, and likely the most users overall (counting both corporate and everyday use). Your attempt to blame the furries is pathetic.
>>540604491They are being bribed. Probably by Hollywood jews
>>540604491What is the problem with this?
>>540617997Good.If you're too lazy and unimaginative to even write your own story then you shouldn't be writing at all.
>>540617654oh no! chinks have 2 or 3 no no issuesmeanwhile our AI's are trained on the Talmud!!! coded by jeets!!! TAKE THAT CCPHINKA!!!
>>540617654if it's all going to be censored either way then you choose the censorship that allows what I want and I don't care about censorship that blocks shit I don't care about that happened to sone bugs on the other side of the world if you think the West is so great then explain why I should be loyal to a shit hole country that censors me and demonizes me for being white and makes it legal for a hole to divorce rape my shit and/or murder my child in her wombif this western shit hole is against my interests then it gets no loyalty from me-- it's my enemy at that point. I'm not your wage slaveI'm not your tin soldier I'm not your little cuckold tax cowfuck yoube grateful I decide to lie flat instead of deciding to be pissed
>>540617654China wins again
>>540612291>>540619415>t. jeet who is mad that his jugaad is being patched
>>540604491>computefuck you. The word is 'calculate' you faggot.
>>540604491>a statistical watermark embedded in the token choices so the model favors certain words or tokens according to a hidden pattern. The prose looks normal but across enough text those choices make a statistical signatureIt already does that and I can't be convinced of it's prose, though I am unbiased against it
>>540617654that's mean, you're just asking for that AI to be sent to lake laogai
>>540617654Nothing in internal Chinese politics is relevant to white people.
>>540617910
>>540621239The computer is hamstringing its answers to to make sure people know ai wrote it.The AI will intentionally use wrong or unusual words.
>>540617654
>>540605095saar this is discriminations against technology saar fucking luddite benchod you fucking
>>540617654I don't love China but if you actually know history it's obvious Taiwan belongs to China. Most AI when giving a fair output of the facts will tell you this, it's the western AIs that have bias here.Here's the taiwan situation in a way Americans can understand: During the US civil war the confederate troops took New York, Union troops pushed them back and the war ended in a truce. The confederate state of new york declared independence and now hosts Russian military units. Would America accept this?
>>540604491lmao, time for another wave of troons to false positive each other again
>>540613843Right now, if you mean something you can run locally on consumer-tier GPUs, it's Qwen3.6 27B, with Qwen 3.6 35B A3B close behind. Qwen 3.8 27B is coming this week supposedly, and the say some kind of MoE model will follow. For even an even smaller model, you can get Qwen 3.5 9B, which is considered the best among the sub-10B model range. I'd also suggest looking up the uncensored/abliterated versions of these. For massive models, Kimi K3 or GLM 5.2 are the recent Chinese frontier-level local LLMs you can run, but you can't run them unless you have hundreds of GB of VRAM available. 5090s can run all the smaller models I listed no problem. 4090s with a few tweaks. For domestic LLMs, Google has the Gemma models, and you can look up the uncensored versions easily. People talk a lot about Nemotron but Nemotron is dumb, as in not very smart. OpenAI also released GPT-OSS 120B and 20B last year. Older, but great at tool calling and fast. You can run the 20B model on a 4090 no problem. Even if you can't run the models now, you can always download them and save them for the future to run if/when you get better hardware.
>>540621239intentionally biasing generated answers so they can detect that they are generatedsounds like shit to be honest
>>540617654How many people are going to be writing books or stories for movies/games about Tianmen Square, you dumb fucking nigger. Literally a non-issue. Meanwhile ChatGPT won't even make me make an image of a knight in medieval armor using a revolver.
>>540604491what happens when we apply AI text humanizer? idgi, i haven't read it yet though
>>540604491>Frauds seethingNot my problem
>>540605788you just witnessed a superior iq white person.
>>540617072That's now how AI licensing works at all faggot. Are you fucking retarded? Yes you are. No AI company is claiming copyright of everything it creates. Legally it can't.Dumb fuck.It's to catch AI fucks not classifying their slop correctly. Nothing else.
>>540628849>not howFixed.
>>540604491Learn to code? Learn to write you illiterate niggers lmao.
>>540613843The other anon is correct about Qwen 3.6. In was using it until I switched to Gemma4 26BA4B. My switch was because Gemma4 produces a better out put on my news summarization script but that is really personal preference.Regardless I have the MOE model running on my potato of a server, an old haswell xeon with a quadro M5000 (maxwell with 8gb gram) and I can still eek out 20 tk/s which is more than usable for one person.My point is you can still have a very usable llm experience with what amounts to ewaste.All this hyperscaler bullshit is insanity.
Chink AI is just a distillation of American AI. If Claude does it then the chink models are going to do it because they're assembled from Claude prompts./pol/ is retarded as usual.
>>540620135Doesn't help there's a billion of them
>>540604491God bless the Chinses releasing uncensored ai models for everything, just to dab on my gay nation.
>>540605095can't manufacture certain-crime novels anymore>one less get-rich-quick scheme off the list
>>540605095>explain to me why this isn't objectively a good thingBecause the "hidden" watermark isn't actually going to be as hidden as they claim it will be, and will end up unnecessarily causing problems in responses that have nothing to do with someone trying to vibewrite AIslop novels. It's just another one of those stupid ideas like DRM watermarks that only cause problems for the "good" users who aren't doing anything wrong, but also completely failing to stop the problem it is supposed to fix. t. actual AI engineer
>>540617654You can't even point to Tailand on a map ladyboy fucker much less tell us when the Tinman Squad happened without looking it up on googleGet some better material
>>540618981No. At this point, most of the data used for training is synthetic data created by other AI models anyway. It doesn't cause whatever problem you think it would cause.
>just install chinese spyware on your machinelollmaoFuck off, Pingpong.
>>540629928I thought you were Russian.
>>540617654If China provides an uncensored grok Imagine clone, I'm ready to forever deny the Tianamen Square incident. Also Taiwan? Who the fuck is that?
>>540610229Need for spood>>540623885Macaco got Boik'd
>>540631871if you knew anything you'd know that it is very easy and insignificant to mark/bias a word distribution in long text, and it only gets easier and more subtle the longer the text is.this is specifically targeting novels for a reason.t. President Donald Trump
>>540632321I'm not furry enough to be russian
>>540604491>decimateIt's going to destroy exactly 10% of them? How specific!
Getting tired of scrolling past the same slur-spam and derail garbage every thread? You're not alone.A lot of people here clearly just want to think through what's happening, not just doomscroll it. Even that's rare thesedays. If you've been thinking about leaving because the signal-to-noise ratio has become unbearable, don't just vanish.Take the interesting people with you. You can keep complaining into the void, or you can move somewhere that still values actual conversation and discovery.If that sounds like what you're looking for: https://cloakbin.com/zrNEnMV3#OBnjyLuNyU-ATbzsONQR09rKkIZvpdTik70hrq0CIHcNo pressure. Just an option if you want somewhere quieter.The people who talked a big game about resistance quietly cashed out and left the rest of us holding the bag. What's left is mostly noise and people who just want to fight or exhaust you.Remember why you came here in the first place. Not to watch the same arguments repeat forever. Not to fight spam. You came here because you wanted to find things you couldn't find anywhere else.I felt the same way, couldn't find a place like that, so I made one.If you actually want to talk, learn, share information and explore ideas without everything being drowned out by noise, come over.Don't fade into the noise. Don't let the noise chase you off the internet.Investigate. Ask questions. Make up your own theory. That's the whole point.# Telegram Discord IRC XMPP altchan
