OpenAI's "unspeakable secrets" have been made public? The method of using YouTube data to train models for acquisition is not glamorous
忧郁草民乜
发表于 2024-3-19 21:47:28
1223
0
0
As the GPT big model continues to advance, OpenAI seems to face only a lot of criticism. Apart from Musk's repeated questioning of "open source", the data source for training the big model in OpenAI is unclear, or it has laid a mine for infringement litigation for this company.
The training of current AI large models requires large-scale datasets with huge amounts of data to be fed on the data side. According to public information, the data sources of OpenAI may include but are not limited to: publicly available data sets, such as various resources on the Internet, such as books, web pages, news articles, academic papers, etc; Datasets provided by partners and third-party data providers; Specific field data purchased, such as medical, legal, or scientific literature; Synthesize data, OpenAI may use its own model to generate data, such as training and improving the model through its own output; Crowdsourcing and community contribution data.
The source of data is not the most important issue, and the focus of external attention is how OpenAI obtains this data.
Did you steal it?
As reported by Business Insider, OpenAI's use of a large number of YouTube videos to train models has become an "open secret", benefiting products including its newly launched Sora model in the field of cultural and biological videos. The mystery lies in how OpenAI can obtain sufficient YouTube content.
You should know that YouTube is a subsidiary of Google. In 2006, YouTube was acquired by Google for $1.65 billion and rapidly grew into the world's largest video sharing platform with Google's support.
Google has always been committed to developing AI and is one of OpenAI's main competitors. Naturally, it will not provide its own gold mine to users for free. YouTube has long banned downloading for commercial purposes and will also restrict the downloading of a large amount of YouTube video data. Under such strict control, individual users have also been affected. Some have stated that even downloading a YouTube video can be very slow, taking several hours to complete.
A common guess is that OpenAI used web crawlers to "steal" YouTube data. OpenAI has admitted to launching a web crawler robot called GPTBot, which is used to crawl and collect data for large-scale model training.
OpenAI executives are evasive about related issues, which indirectly deepens people's impression of it as a "data thief". The Wall Street Journal recently asked OpenAI Chief Technology Officer Mira Murati if the startup is using videos from YouTube, Instagram, and Facebook to train Sora.
"I'm actually not sure about this," she said. When asked again about the source of the training data, Murati refused to answer, "I won't disclose any details."
According to a recent report from Business Insider, a person familiar with OpenAI operations stated that the company has assigned a tightly protected team to obtain training data, and how to obtain this data is a confidentiality issue.
The AI field with dense fences
The use of crawlers is not tolerated by Google, as its YouTube platform prohibits the use of robots and other automated methods to crawl its videos.
But for OpenAI, accessing YouTube videos in a way that violates Google's terms of service may not be illegal. The case law and "fair use" principle in the United States grant companies the right to freely use online content in different ways.
In short, Google, OpenAI, and other technology companies currently believe that using copyrighted content for artificial intelligence model training is also legal. Regulatory authorities have not yet made clear regulations on this matter. The arena of artificial intelligence remains a vast wilderness, with game rules related to data either yet to be determined or ignored.
Manufacturers are competing to enter and build their own technological barriers.
OpenAI and other large model developers have previously disclosed their training data sources in published research papers, but this practice is no longer prevalent as competition intensifies. Everyone wants to retain their technical secrets in order to gain a relative advantage, especially for top manufacturers who hold a favorable position. The open source competition is also a manifestation of manufacturers trying to keep their own trump cards.
The only certainty is that with further iteration of generative AI technology, similar disputes will only increase.
Big companies are more likely to become targets of criticism. Taking data as an example, even if they dare to take responsibility and bear high data procurement costs, achieving complete compliance in data acquisition is not easy. Due to the large number of parameters, large models require the use of distributed computing and cloud services for training and deployment, which increases the risk of data theft, tampering, abuse, or leakage.
How to balance personal privacy protection and encourage technological innovation, and how to find the optimal path between enterprise survival and compliant production, has become an unavoidable issue for every company committed to generative AI.
CandyLake.com 系信息发布平台,仅提供信息存储空间服务。
声明:该文观点仅代表作者本人,本文不代表CandyLake.com立场,且不构成建议,请谨慎对待。
声明:该文观点仅代表作者本人,本文不代表CandyLake.com立场,且不构成建议,请谨慎对待。
猜你喜欢
- Apple lowers prices of various iPhone models in India
- Baidu Shen Dou: Upgrade computing platform capability for 100000 card computing power cluster, Wenxin large model daily usage exceeds 700 million times
- Meta releases heavyweight new products: $299 Quest 3S headset, AR glasses prototype, multimodal AI model
- Baidu World 2024 will be held on November 12th, and the daily average number of adjustments for the Wenxin large model has exceeded 700 million times
- 挑战Model Y 蔚来的品牌下沉“阳谋”
- Ford CEO tired of making 'boring' car models, personalized and electrified products become 'new favorites'
- Dialogue | Baidu Li Tao: The overlap between automotive intelligence and the wave of big models is a historical inevitability
- Boeing announces 10% layoffs, first delivery of 777X model postponed to 2026
- Faraday Future plans to launch the first model of its second brand by the end of next year
- Will a third brand launch hybrid models overseas? NIO responds: Continuing the pure electric technology route
-
アップル(AAPL)は現地時間10月31日、9月28日までの2024年第4四半期の業績を発表し、同四半期の売上高は前年同期比6%増の949億3000万ドル、純利益は同36%減の147億3600万ドルだった。 純利益の下落は主に102億ドル ...
- 亲亲宝宝贝贝贾
- 3 天前
- 支持
- 反对
- 回复
- 收藏
-
【英偉達の需要が高すぎる?SKハイニックス:黄仁勲がHBM 4チップの6カ月前納入を要求!】SKハイニックスの崔泰源(チェ・テウォン)会長は月曜日、インビダーの黄仁勲(ファン・インフン)CEOが同社の次世代高帯域 ...
- 琳271
- 昨天 17:54
- 支持
- 反对
- 回复
- 收藏
-
ファイザーが前立腺がんを治療する革新薬テゼナ& ;reg;(TALZENNA®,一般名:トルエンスルホン酸タラゾールパーリカプセル)は2024年10月29日に国家薬品監督管理局(NMPA)の承認を得て、HRR遺伝子突然変異 ...
- 什么大师特
- 1 小时前
- 支持
- 反对
- 回复
- 收藏
-
南方財経は11月5日、中央テレビのニュースによると、現地時間11月5日、米ボーイング社のストライキ労働者が59%の投票結果で新たな賃金協定を受け入れ、7週間にわたるストライキを終えた。ストライキ労働者は11月12 ...
- Dubssgshbsbdhd
- 2 小时前
- 支持
- 反对
- 回复
- 收藏