Apple Intelligence was not trained on the free database The Pile, which contains subtitles from thousands of YouTube videos without asking their creators. The company announced this on the Apple blog 9to5Mac. The company wrote in a scientific paper on its high-efficiency models in the OpenELM series that the data set was being used. However, OpenELM is just that Is not part of the AI system used by the companyThat includes Apple Intelligence or other machine learning technologies.
Advertisement
Subtitles from 170,000 YouTube videos
According to 9to5Mac Apple said it developed OpenELM as a contribution to AI research and the advancement of open source language models. At the time, the company described the technology as a “cutting edge open language model.” OpenELM was developed for research purposes only, not to operate any Apple intelligence function. OpenELM is still there On Apple’s AI research website Available.
The training data set “The Pile” was criticized, which comes from the non-profit organization EleutherAI In a report by The Proof According to which other big companies like Nvidia, Anthropic and Salesforce also use the information. Among other things, “The Pile” is considered 170,000 YouTube videos with subtitles It has been fed. It is said that no approval has been received for this.
Question about training data
It is still unclear what and how much training data Apple uses for Apple Intelligence. The company only says that it “uses licensed content, including data that improves specific functions.” However, there is also data that Apple itself has obtained from the public Internet with its own web crawler.
To opt out, website operators must instruct the special “Applebot Extended” to ignore their content. Crawling of websites by AppleBot (which is not used for AI purposes, but for other services) continues even after opting out, if it is not denied at the same time in the “robots.txt” file, writes Company at Apple.com. It is also known that the group does not involve users’ personal data or “user interactions” in training. There are also filters for credit card information or “obscenity” and low-quality content – although it is unclear how these are excluded.
(B.Sc.)
