Anthropic is now watermarking everything Claude writes. But how do you hide a watermark in plain text?
In this video, I break down how LLM watermarking actually works, starting with the red/green token lists of Kirchenbauer et al. and ending with the tournament sampling behind Google DeepMind’s SynthID-Text, which is the method Anthropic says it will use for Claude. I also build all three schemes in a playground so you can see exactly what they do to the output.
☕ Support the channel: https://www.patreon.com/NoHypeAI
🎥 CHAPTERS
──────────────────
00:00 – Anthropic Is Watermarking Claude
01:15 – The Hard Red List
03:12 – Seeds, Hashes and the Detector
05:07 – The Problem With Entropy
05:52 – Demo: Hard Red List
07:42 – The Soft Red List
09:58 – Demo: Soft Red List
11:35 – SynthID-Text and Tournament Sampling
15:27 – Demo: Tournament Sampling
16:18 – What This Means in Practice
📖 USEFUL LINKS
──────────────────
👉 Kirchenbauer et al., “A Watermark for Large Language Models” – https://arxiv.org/abs/2301.10226
👉 Dathathri et al., “Scalable watermarking for identifying large language model outputs” (Nature) – https://www.nature.com/articles/s41586-024-08025-4
👉 SynthID-Text reference implementation – https://github.com/google-deepmind/synthid-text
👉 Anthropic’s watermarking announcement page – https://www.anthropic.com/news/claude-text-watermark
#AI #LLM #Watermarking #Claude #MachineLearning
source
Leave a Reply