AI - AI Dynamics

→ View original post on X — @packym

Samsung NFL Ads Disappear: Mystery of Missing App Advertisements

By

@packym

–

25 January 2023 23h24

Don’t tell Mr. Tweet, but I haven’t been served an ad on this app in months, maybe years. One day, Samsung or somebody was promoting two-week old NFL highlights into my feed, then… nothing.

→ View original post on X — @googleai

ML Models Improve Multi-Armed Bandit Algorithm Performance Online

By

@googleai

–

25 January 2023 23h09

Existing algorithms for the multi-armed bandit problem do not account for the available real world data that can aid algorithm design. Learn how an ML model that provides a weak hint can improve the performance of an algorithm in an online setting → https://
goo.gle/3XF84b0

Watermarking: a necessary trick for AI-generated content

By

–

25 January 2023 22h54

Watermarking will be a necessary trick to sift through the chaos of AI-generated content in the future. Authors: John Kirchenbauer, Jonas Geiping, Yuxin Wen Jonathan Katz, Ian Miers, Tom Goldstein from University of Maryland Link to paper: https://
arxiv.org/pdf/2301.10226
.pdf
…

AI systems alerted to avoid synthetic data to prevent recursive training problem

By

–

25 January 2023 22h54

More importantly, it can alert the AI systems themselves not to train on "synthetic data", which could help solve the recursive AI data problem (AI training itself on it's own output, AKA the machine learning version of eating your own vomit)

Watermark tokens embedded with negligible impact on text quality.

By

–

25 January 2023 22h54

Interestingly, the watermark tokens can be embedded with negligible impact on text quality.

Watermark selects whitelist tokens and promotes them during sampling

By

–

25 January 2023 22h54

The watermark works by selecting a randomized set of “whitelist” tokens before any text is generated, and then softly promoting use of those tokens during sampling. (a token can be a letter, word, numbers, punctuations or any other string that is useful for the specific task)

LLM watermark detection: 28 tokens vs expected 9, odds 6e-14

By

–

25 January 2023 22h54

In this sample from the paper, text written by a human would likely contain a certain number of "whitelisted" tokens (expected to be about 9 in this sample), but the text produced by the LLM with the watermark contains 28. Odds of this happening randomly = 6 × 10^−14.