They actually run 2^30-way tournament (they derive an equivalent form that doesn... | Hacker News

Hacker Newsnew | past | comments | ask | show | jobs | submit

		mmoskal 9 months ago \| parent \| context \| favorite \| on: DeepMind debuts watermarks for AI-generated text They actually run 2^30-way tournament (they derive an equivalent form that doesn't requires 2B operations). You do not need to run the LLM, it only depends on the tokenizer.

espadrine 9 months ago [–]

You’re right. I understood it to require taking the top 2^30 tokens, but instead they sample 2^30 times with replacement.

Too bad they only formulate the detection positive rate empirically. I am curious what the exact probability would be mathematically.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact