In a legal dispute before a California court, attorneys for several prominent US authors accuse Meta CEO Mark Zuckerberg of personally approving the use of pirated content for training Meta’s AI models. According to recently released court documents, the AI team reportedly received permission to use data from LibGen for training the Llama models after “escalating to MZ.” Initially, Meta developers expressed concerns about using the data on company laptops. Meta is also accused of deliberately removing copyright notices from e-books and scientific articles from LibGen, presumably to avoid indications of copyrighted material in AI responses. Developers even had to upload protected material themselves to gain access to file-sharing platforms.
In the ongoing legal case (Kadrey et al. v. Meta Platforms, Inc.), US authors Sarah Silverman, Richard Kadrey, and Christopher Golden are suing Meta for the illegal use of their books for AI training. While the court dismissed most allegations in September 2023, the accusation of copyright infringement through AI training remains. Meta has not yet responded to inquiries about this issue.
OpenAI, Google, and other AI companies have found a new way to acquire training material: they buy unpublished video content directly from content creators. The companies pay between one and four dollars per minute, with high-quality 4K footage and special shots like drone videos or 3D animations in high demand. Ordinary, unused material from YouTube, Instagram, or TikTok is paid one to two dollars per minute.
The companies work with specialized licensing firms like Troveo AI and Calliope Networks. According to Troveo, they have already paid over five million dollars to creators. Marty Pesis, co-founder and CEO of Troveo, told Bloomberg that all companies developing video models are either already working with them or are in the pipeline. Dan Levitt from the Wasserman agency sees a competitive race for video material among companies. He predicts a window for lucrative licensing opportunities in the coming years, but it won’t stay open forever.
Elon Musk’s AI startup xAI has released its Grok app for iOS in the US. The free app is based on the Grok 2 language model and allows for questions, image generation, and photo analysis. Grok uses current Twitter and web data for its responses. The app is currently only available in the US for iOS. The announced Grok 3 model is reportedly in beta testing and is expected to be released in the coming weeks.
AMD and Johns Hopkins University have developed a new AI framework called Agent Laboratory. This open-source tool is designed to support scientists in their research work by combining human creativity with AI-driven workflows. Agent Laboratory operates in three main phases: a PhD agent first conducts a literature review on a human-specified topic and collects relevant research papers via the arXiv API. Then, PhD and postdoc agents create a detailed research plan. An ML-engineer agent implements and conducts the experiments. Finally, the agents summarize the results in an academic report.
In tests, the o1-preview model within the framework performed best in terms of clarity and relevance. The cost per research paper was nearly 13 US dollars with o1-preview, while it was just over 2 US dollars with GPT-4o. Weaknesses included inaccurate self-assessments by the AI and the risk of hallucinations. Additionally, machine evaluations overestimated the quality compared to human reviewers.
Stability AI has introduced the SPAR3D model at CES in collaboration with Nvidia. This 3D reconstruction tool can generate complete 3D objects from a single image in less than a second. SPAR3D also allows for direct editing of the generated point clouds, which can be converted into finished 3D meshes in just 0.3 seconds. The system is designed for PCs with Nvidia RTX graphics cards. The model is available for free under the Stability AI Community License. Only organizations with annual revenues over one million dollars require an enterprise license.