Non-profit AI analysis group EleutherAI scraped YouTube subtitles to create a dataset in violation of YouTube’s phrases of service, ProofNews stated on July 16.
The dataset, known as the Pile, allegedly consists of subtitles of 173,536 YouTube movies from over 48,000 channels. About 12,000 deleted movies are a part of the dataset.
A number of high tech and AI companies, together with Anthropic, have since used the Pile for coaching. Anthropic spokesperson Jennifer Martinez stated the dataset consists of “a really small subset of YouTube subtitles” however declined to touch upon attainable violations of YouTube’s phrases of service.
Enterprise software program agency Salesforce additionally used the dataset. Salesforce VP of AI analysis Caiming Xiong stated the dataset was “publicly out there” and that Salesforce used it for tutorial and analysis functions. ProofNews stated Salesforce finally launched the identical dataset publicly.
Apple used the Pile to coach OpenELM, an environment friendly language mannequin for on-device AI. Nvidia, Bloomberg, and Databricks additionally used the Pile for AI coaching.
ProofNews stated its listing of corporations that used the dataset shouldn’t be complete, as corporations don’t all the time disclose which datasets they use in AI coaching.
Dataset incorporates crypto channels, extra
ProofNews’ search software signifies that Pile consists of movies from crypto channels and creators, together with Coinbase, Cointelegraph, Bitcoin Journal, BitBoy Crypto, 99Bitcoins, Ivan On Tech, and Andreas Antonopolous.
ProofNews highlighted that the dataset consists of transcripts from main information channels, training channels, late-night exhibits, in style YouTube hosts, and different classes. The Pile dataset extends past YouTube to different web sites and on-line content material.
ProofNews famous an earlier report from the New York Occasions, which stated OpenAI and Google had beforehand harvested YouTube textual content. Google, which owns YouTube, stated the motion was permissible attributable to its settlement with customers. OpenAI didn’t verify or deny the report.
AI copyright disputes are far-reaching. Regulation agency Baker Hoestler lists a minimum of fifteen lawsuits involving tech companies similar to Anthropic, Meta, GitHub, Stability AI, Nvidia, and Google. OpenAI faces high-profile lawsuits from Mom Jones’ mother or father firm and The New York Occasions.