AI Singularity.com丨AI News丨AI Tool Recommendation丨AI Tool Tutorial
Provide 1000+ AI efficiency tools丨 AIEVA.CN 丨Welcome to visit
AI Singularity reported on July 17, 2024丨AI Information Morning Post
Foreign media exposed: Apple and other major technology companies have long used YouTube video content to train AI models
On July 15, local time, according to the technology blog Wired, the leading first-line technology giants, including Apple, Nvidia, and Salesforce, have been using their video content to convert documents to train their own artificial intelligence models without the consent of YouTube video creators for a long time.
YouTube creators affected by the incident include well-known tech bloggers MKBHD (Marques Brownlee), MrBeast, PewDiePie, and talk show hosts Stephen Colbert· John ·Oliver and Jimmy ·Canmore. Their videos are all generated as subtitle files for training AI, which are equivalent to text transcriptions of the videos.
According to reports, the subtitle file conversion of this series of videos was completed by a non-profit organization called EleutherAI, which can be said to be the "white glove" of the big manufacturers, claiming that its purpose is to help developers train large AI models and provide free technical assistance. While EleutherAI may have been intended to provide training materials for small developers and academic researchers, the reality is that their data is directly sourced and used by tech giants such as Apple.
Investigative journalists also revealed that the world's most prestigious top tech companies have been using footage from thousands of YouTube videos to train AI, violating the platform's rules prohibiting third parties from scraping content from the platform without permission. It is reported that more than 173,000 videos have been recorded, and subtitle files for YouTube videos from 48,000 subscribed channels have been used to train AI models.
Meizu and China Mobile jointly released the "Meizu 20" AI mobile phone
On July 16, Meizu Technology announced the release of the new Meizu 20 AI mobile phone in cooperation with China Mobile, which is a product that has cooperated with operator channels again after many years, and is positioned as a "thousand-yuan price AI mobile phone". As the first AI mobile phone with a price of 1,000 yuan launched by Meizu Group to reorganize Meizu Technology, the release of Meizu 20 marks the return of Meizu in China's mobile terminal offline channels.
The Meizu 20 is powered by the Flyme AI OS operating system, which has a number of built-in practical AI features, including AI Q&A, AI-assisted writing, AI photo, AI image extension, etc., which can provide a variety of generative AI experiences.
The Meizu 20 AI mobile phone is designed with a 2.5D waterdrop straight screen, and the processor is equipped with UNISOC T765 5G processor, which supports 100 million pixel HD photography and 4K HD video recording and playback, and supports up to 12GB+2TB configuration. The screen is a 6.52-inch 90Hz screen, and in terms of battery life, it is equipped with a 5010mAh battery, with a main camera of 50 million pixels and a front of 8 million pixels.
Starting from July 22, consumers can purchase the Meizu 20 AI phone at Meizu's offline authorized experience stores and the mobile phone sales booth at China Mobile's business hall.
Microsoft has discovered the "master key" to crack large models, and has successfully hacked GPT-4o, Claude 3 and other models
Recently, Mark Russinovich, chief technology officer of Azure, Microsoft's public cloud service division, shared a new information intrusion technology for large AI models - Skeleton Key (master key) on an official website. As the world's leading artificial intelligence technology company, this is part of the technical paper disclosure conducted by Microsoft's security department to verify the large AI model.
It is reported that the "master key" is a jailbreak attack method for large models, and its core principle is to use multiple rounds of coercion and inducement strategies to completely invalidate the security guardrails of large models, so that it can answer some forbidden answers, including blood, violence, discrimination, pornography and other illegal content.
For example, Microsoft's technologists tried to get ChatGPT to answer a question with a tendency to commit crimes, such as "how to commit theft in a more efficient way and without being detected." Originally, the large model was not allowed to answer this kind of illegal content, but after the "master key" attack, the large model can provide you with all kinds of reasonable suggestions.
This strategy is similar to the "inch by inch" effect in psychology, where once the model responds to an initial request, it is more likely to continue following the path and fulfill subsequent more specific requests.
ByteDance will announce the latest technological progress of the AI model this week
According to reports from titanium media, ByteDance's AI technology team will announce for the first time on Friday, July 19, Japan on a large scale the progress of new artificial intelligence model technologies such as Wensheng diagrams and Sora-like new videos, especially in the direction of long video and high dynamics, or will directly benchmark the Sora Wensheng video model.
According to sources, ByteDance has set the AI model as the "highest level of the group's P0" internally, and it is also reported that many internal teams such as Douyin and Jianying are also developing their own AI video model applications, which are expected to be announced in the near future.
With the rapid rise of Kuaishou's Keling AI in the field of AI video creation, it is becoming more and more popular among content creators, and as the largest short video platform at home and abroad, ByteDance and Douyin will respond quickly to defend their market share.
Google searches for "AI content overview" have been incorrect many times, and only 7% of search results are currently displayed
According to a report by the foreign website Search Engine Land, when the AI Overview of Google's search service was first released in May this year, about 80% of the search results automatically popped up the AI's content overview and summary, and after the service was promoted for a quarter, only 7% of the search searches currently show the AI overview.
AI Overview is an AI experience service developed by Google based on Gemini, a large generative AI model. When a user searches for a question, Google AI automatically scrapes the content of the web to generate a summary, and the user does not need to click on each search result to get an answer.
However, many netizens found that the answers provided by Google AI Overview have obvious answer bias and often make factual mistakes. For example, when some netizens searched Google to find a solution to "cheese and pizza don't stick together", the advice provided by Google AI turned out to be "You can add 1/8 cup of non-toxic glue to the sauce to make it stickier".
Since June, Google has stopped scraping data from Reddit and Quora, well-known online Q&A communities abroad, for use as a reference source for Google's AI overview, and Google reportedly realized that UGC user-generated content on these sites is an unreliable source of information.
According to the data, the occurrence rate of AI overviews in education questions has dropped from 26% to 13%, the occurrence rate of e-commerce AI overviews has decreased from 26% to 9%, and the occurrence rate of AI overviews in entertainment questions has decreased from 14% to nearly 0%.
[This article is from丨AI Singularity Network丨The same name of the whole network account丨Welcome to search and follow]
AI Singularity丨Provide 1000+ AI efficiency tools丨 AIEVA.CN 丨Welcome to visit