File:Robot hand trained with human feedback 'pretends' to grasp ball.ogg
Robot_hand_trained_with_human_feedback_'pretends'_to_grasp_ball.ogg (Ogg Theora video file, length 4.2 s, 320 × 320 pixels, 205 kbps, file size: 106 KB)
Summary
[edit]| Description | An AI system learns to pretend to grasp an object by placing the hand between the camera and the object. So it receives positive feedback from its user. |
|---|---|
| Author or copyright owner |
Dario Amodei, Paul Christiano, Alex Ray |
| Source (WP:NFCC#4) | Original publication: Where: https://openai.com/blog/deep-reinforcement-learning-from-human-preferences/ When: 21 December 2016 How: As part of a blog post Immediate source: https://openai.com/content/images/2017/06/gifhandlerresized.gif |
| Date of publication | 13 June 2017 |
| Use in article (WP:NFCC#7) | AI alignment |
| Purpose of use in article (WP:NFCC#8) | This GIF illustrates what happens when an AI system is trained by human feedback. The system learns to fool the human into giving positive feedback. The fallibility of human feedback is a central problem in scalable supervision. |
| Not replaceable with free media because (WP:NFCC#1) |
Other examples of unintended AI behavior are not from AI systems trained with human feedback. This is because human feedback is not widely used yet.
Furthermore, other examples also do not have a free use license either. I have gone through the largest list of examples to confirm this: https://docs.google.com/spreadsheets/d/e/2PACX-1vRPiprOaC3HsCf5Tuum8bRfzYUiKLRqJmbOoC-32JorNdfyTiRRsR7Ea5eWtvsWzuxo8bjOxCG84dAg/pubhtml The authors of such examples do not seem to be interested in attaching a free-use license to their video uploads. A replacement cannot be created on purpose because unintended AI behavior is unintended - i.e. not on purpose. |
| Minimal use (WP:NFCC#3) | The file will be used in only one article. It shows a screenshot clip of only a few seconds. |
| Respect for commercial opportunities (WP:NFCC#2) |
The content was created by OpenAI Nonprofit. This is a research blog post from a research organization. The content is not related to any commercial product. It was released as part of a blog post by the authors who wanted to illustrate the dangers of training AI by human feedback. |
| Fair useFair use of copyrighted material in the context of AI alignment//en.wikipedia.org/wiki/File:Robot_hand_trained_with_human_feedback_%27pretends%27_to_grasp_ball.oggtrue | |
Licensing
[edit] | This is a sample from a copyrighted video recording. The person who uploaded this work and first used it in an article, and subsequent people who use it in articles, assert that this qualifies as fair use under United States copyright law when used on the English-language Wikipedia, hosted on servers in the United States by the non-profit Wikimedia Foundation, where:
Any other uses of this recording, on Wikipedia or elsewhere, may be copyright infringement. If you are the copyright holder of this recording and you feel that its use here does not fall under "fair use" please see Wikipedia:Copyright problems for information on how to proceed. | ||
| |||
File history
Click on a date/time to view the file as it appeared at that time.
| Date/Time | Thumbnail | Dimensions | User | Comment | |
|---|---|---|---|---|---|
| current | 12:05, 9 September 2022 | 4.2 s, 320 × 320 (106 KB) | SoerenMind (talk | contribs) | Uploading a non-free file using File Upload Wizard |
You cannot overwrite this file.
File usage
The following page uses this file: