The more apt word is steganography, rather than watermark. Basically subtly adjust the weights of the model so that some subtle patterns appear. Think of how AI text prefers certain words and phrases that ordinary humans don’t use as often, like “delve,” but presumably much more subtle.
And no, as Anthropic has already said, this watermark may not survive editing/formatting.
Claude Now Watermarks Your Text | Vanja Petreski - vanja.io/claude-invisible-watermark/
Replying to @PetteriPano@lemmy.world
Chinese being more token efficient is a myth, and seems to stem from the superficial fact that characters are only visually more space efficient.
The fact that each Chinese character takes up 3 bytes (as opposed to 1 byte of English), words in Chinese typically require compounds of several characters, and that tokenizers have a limited vocabulary limited to mostly English means that Chinese is actually token inefficient.
No, Chinese Is Not More Token-Efficient Than English for LLMs | markhuang.ai - markhuang.ai/blog/chinese-token-myth
markhuang.aiNo, Chinese Is Not More Token-Efficient Than English for LLMsA native Mandarin speaker tests the popular claim that Chinese characters save tokens when interacting with LLMs. Across six tokenizers — including Chinese-first models like Qwen, GLM, and DeepSeek — English uses fewer tokens every time. The data, the BPE mechanics, and why character count has nothing to do with token count.Replying to @712@discuss.tchncs.de
Nemotron Ultra is a open source and already gives DeepSeek R1 performance (admittedly not that good anymore). NVIDIA has open sourced the entire training process, including raw datasets and synthetic data generation. The raw data is mostly curated web crawl data.