Timeline for GPT on the platform: Data, actions, and outcomes
Current License: CC BY-SA 4.0
14 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Jun 9, 2023 at 11:49 | comment | added | Wrzlprmft | @Trilarion: And why not plotting the average Huggingface score of all new answers per week for the last 12 months including deleted posts for example. – I guess Hugging Face and similar have rate limits for their services, which is why SE’s study only used 500 posts. If they could easily analyse all posts in the year before December 2022, they probably would have done so. | |
| Jun 9, 2023 at 11:47 | history | edited | Wrzlprmft | CC BY-SA 4.0 |
added 162 characters in body
|
| Jun 9, 2023 at 11:38 | comment | added | NoDataDumpNoContribution | I'm just trying to find alternative ways to estimate the amount of GPT contributions or as a sanity check. Short of requiring additional webcam live feeds for every question and answer, all reasonable information should be used. And why not plotting the average Huggingface score of all new answers per week for the last 12 months including deleted posts for example. If then there is a statistical significant change in Dec. we know what happened and the direction of the time evolution of the score might also tell us if the problem got worse or better in the mean time. | |
| Jun 9, 2023 at 11:16 | comment | added | Wrzlprmft | While a similar problem exists with the analysis of saved drafts, it can at least directly focus on all relevant posts. | |
| Jun 9, 2023 at 11:15 | comment | added | Wrzlprmft | @Trilarion: The problem with this approach is that the fraction of interesting posts (GPT-produced stuff) is very small, so you would have to run a lot of posts through Hugging Face and any trends will have a strong influence. It’s like trying to determine the number of five-person households with a car by subtracting the fraction of households with a car and four or fewer persons from the fraction of all households for a car, but measured in a different year. It’s essentially catastrophic cancellation, with some complexity on top. | |
| Jun 9, 2023 at 10:50 | comment | added | NoDataDumpNoContribution | Maybe one could run this detector on old pre2023 data, on 2023 data and on purely GPT produced data, make a histogram of the scores and try to approximate the 2023 histogram as combination of both others. | |
| Jun 8, 2023 at 20:24 | comment | added | Wrzlprmft | @This: It’s a rate of false positives not a positive rate that is false. | |
| Jun 8, 2023 at 20:23 | history | edited | Wrzlprmft | CC BY-SA 4.0 |
It’s a rate of false positives not a positive rate that is false.
|
| Jun 8, 2023 at 17:11 | history | edited | This_is_NOT_a_forum | CC BY-SA 4.0 |
Second iteration [<https://en.wikipedia.org/wiki/Hugging_Face>].
|
| Jun 8, 2023 at 17:04 | history | edited | This_is_NOT_a_forum | CC BY-SA 4.0 |
Active reading [<https://en.wikipedia.org/wiki/Hugging_Face> <https://en.wiktionary.org/wiki/in_other_words#Prepositional_phrase> <https://en.wikipedia.org/wiki/True_negative> <https://en.wiktionary.org/wiki/false_positive#Noun>].
|
| Jun 7, 2023 at 21:49 | history | edited | Wrzlprmft | CC BY-SA 4.0 |
edited body
|
| Jun 7, 2023 at 21:43 | history | edited | Wrzlprmft | CC BY-SA 4.0 |
added 13 characters in body
|
| Jun 7, 2023 at 21:33 | history | edited | Wrzlprmft | CC BY-SA 4.0 |
added 143 characters in body
|
| Jun 7, 2023 at 21:13 | history | answered | Wrzlprmft | CC BY-SA 4.0 |