brainsteam.co.uk/brainsteam/content/annotations/2022/12/19/1671461752.md

date	hypothesis-meta	in-reply-to	tags	type	url
2022-12-19T14:55:52	created document flagged group hidden id links permissions tags target text updated uri user user_info 2022-12-19T14:55:52.384335+00:00 title My AI Safety Lecture for UT Effective Altruism false __world__ false O7YUan-tEe29vjfmuBFMKQ html incontext json https://hypothes.is/a/O7YUan-tEe29vjfmuBFMKQ https://hyp.is/O7YUan-tEe29vjfmuBFMKQ/scottaaronson.blog/?p=6823 https://hypothes.is/api/annotations/O7YUan-tEe29vjfmuBFMKQ admin delete read update acct:ravenscroftj@hypothes.is acct:ravenscroftj@hypothes.is group:__world__ acct:ravenscroftj@hypothes.is explainability nlproc selector source endContainer endOffset startContainer startOffset type /div[2]/div[2]/div[2]/div[1]/p[95] 193 /div[2]/div[2]/div[2]/div[1]/p[95] 0 RangeSelector end start type 38138 37945 TextPositionSelector exact prefix suffix type So then to watermark, instead of selecting the next token randomly, the idea will be to select it pseudorandomly, using a cryptographic pseudorandom function, whose key is known only to OpenAI. of output tokens) each time. That won’t make any detectable TextQuoteSelector https://scottaaronson.blog/?p=6823 Watermarking by applying cryptographic pseudorandom functions to the model output instead of true random (true pseudo-random) 2022-12-19T14:55:52.384335+00:00 https://scottaaronson.blog/?p=6823 acct:ravenscroftj@hypothes.is display_name James Ravenscroft	https://scottaaronson.blog/?p=6823	explainability nlproc hypothesis	annotation	/annotations/2022/12/19/1671461752

So then to watermark, instead of selecting the next token randomly, the idea will be to select it pseudorandomly, using a cryptographic pseudorandom function, whose key is known only to OpenAI.

Watermarking by applying cryptographic pseudorandom functions to the model output instead of true random (true pseudo-random)