AI

NVIDIA reportedly scraped vast amounts of copyrighted content for AI training. On Monday, 404 Media’s Samantha Cole reported that the $2.4 trillion company instructed workers to download videos from YouTube, Netflix, and other datasets to develop commercial AI projects. The graphics card maker seems to have embraced a “move fast and break things” approach as it competes for dominance in the intense and often controversial AI gold rush. The training was allegedly used to develop models for products like its Omniverse 3D world generator, self-driving car systems, and “digital human” initiatives. YouTube seems to disagree. Spokesperson Jack Malon referred us to a Bloomberg story from April, where CEO Neal Mohan stated that using YouTube to train AI models would be a clear violation” of its terms. NVIDIA employees who raised ethical and legal concerns about the practice were reportedly told by their managers that it had already been approved by the company’s top executives. “This is an executive decision,” responded Ming-Yu Liu, vice president of research at NVIDIA. “We have an umbrella approval for all of the data.” Others in the company allegedly described the scraping as an “open legal issue” to be addressed later. The situation is reminiscent of Facebook’s (now Meta’s) old “move fast and break things” motto, which famously compromised the privacy of millions of users. In addition to YouTube and Netflix videos, NVIDIA reportedly instructed employees to train on other datasets, including the movie trailer database MovieNet, internal libraries of video game footage, and Github video datasets WebVid (now removed after a cease-and-desist) and InternVid-10M, which contains 10 million YouTube video IDs. Some of the data that NVIDIA allegedly trained on was only intended for academic or non-commercial use. For instance, HD-VG-130M, a library of 130 million YouTube videos, includes a usage license specifying it’s for academic research only. NVIDIA reportedly dismissed concerns about these academic-only terms, insisting the data was fair game for its commercial AI projects. To avoid detection by YouTube, NVIDIA reportedly used virtual machines (VMs) with rotating IP addresses to download content and avoid bans. When a worker suggested using a third-party IP address-rotating tool, another NVIDIA employee reportedly responded, “We are on Amazon Web Services and restarting a virtual machine instance gives a new public IP. So, that’s not a problem so far.”

YouTube has introduced a new AI assistant feature to help users recover hacked accounts and protect them from future attacks. The announcement appeared earlier today on Google’s YouTube support page. The new “hacked channel assistant,” available on YouTube, provides “eligible creators” with a way to detect their accounts after being hacked. The feature can be accessed through the YouTube Help Center. The assistant guides affected users through a series of questions to secure their Google login, reverse any changes made by the hacker, and safeguard their channel from further unauthorized access. Currently, the feature is available only in English and for a select group of “certain creators,” but Google is working to make it accessible to all YouTube creators.

Google Meet is introducing a new AI tool called “Take Notes for Me,” which generates summaries of key points during video calls. Instead of providing a word-for-word transcription, this feature uses Gemini AI to capture key discussion points in a Google Doc that will appear in the meeting owner’s Google Drive. The document can be automatically sent to attendees or added to the calendar event after the call, and it will include links to the meeting recording and transcript if those features are enabled. Google Workspace customers with the Gemini Enterprise, Gemini Education Premium, and AI Meetings & Messaging add-ons will be the first to access this note-taking feature, which is currently limited to English and available only for meetings on computers or laptops. Last year, the tech company announced plans to incorporate generative AI into more services and has aggressively integrated AI features into its hardware and software since then. Its Gemini AI assistant is being embedded into Android and Workspace apps. However, the effectiveness of these features will depend on the accuracy and reliability of the AI.

To reflect local values and regulations, nations are increasingly pursuing sovereign AI strategies by developing AI that utilizes their own infrastructure, data, and expertise. NVIDIA supports this trend with the launch of four new NVIDIA Neural Inference Microservices (NIM). These microservices simplify the creation and deployment of generative AI applications, supporting regionally-tailored community models. They promise deeper user engagement by enhancing understanding of local languages and cultural nuances, leading to more accurate and relevant responses. This development comes as the Asia-Pacific generative AI software market is expected to experience a significant boom. ABI Research predicts revenue will surge from $5 billion this year to an impressive $48 billion by 2030. Among the new offerings are two regional language models: Llama-3-Swallow-70B, trained on Japanese data, and Llama-3-Taiwan-70B, optimized for Mandarin. These models are designed to better understand local laws, regulations, and cultural nuances. The RakutenAI 7B model family further strengthens the Japanese language offering. Based on Mistral-7B and trained on English and Japanese datasets, these models are available as two separate NIM microservices for Chat and Instruct functions. Notably, Rakuten’s models achieved the highest average score among open Japanese large language models in the LM Evaluation Harness benchmark from January to March 2024. Training LLMs on regional languages is essential for improving output efficacy. By accurately reflecting cultural and linguistic subtleties, these models enable more precise and nuanced communication. Compared to base models like Llama 3, these regional variants show superior performance in understanding Japanese and Mandarin, handling regional legal tasks, and translating and summarizing text. The global push for sovereign AI infrastructure is reflected in significant investments from countries such as Singapore, UAE, South Korea, Sweden, France, Italy, and India. “LLMs are not mechanical tools that provide uniform benefits; rather, they interact with human culture and creativity. The impact is reciprocal: models are shaped by the data they are trained on, and our culture and data are influenced by LLMs,” said Rio Yokota, professor at the Global Scientific Information and Computing Center at the Tokyo Institute of Technology. “Therefore, developing sovereign AI models that adhere to our cultural norms is crucial. The availability of Llama-3-Swallow as an NVIDIA NIM microservice will allow developers to easily access and deploy the model for Japanese applications across various industries.” NVIDIA’s NIM microservices enable businesses, government bodies, and universities to host native LLMs within their own environments. Developers can create advanced copilots, chatbots, and AI assistants with these services, which are optimized for inference using the open-source NVIDIA TensorRT-LLM library, promising improved performance and deployment speed.

A data privacy controversy involving the social media platform X has placed the European Union in the spotlight. On August 8, an Irish court ruled that X had agreed to suspend the use of all data from European Union citizens that had been collected via the platform for training the company’s AI systems. According to The Economic Times, this decision was driven by complaints from the Data Protection Commission (DPC) of Ireland, which is the primary EU regulator for many large U.S. tech companies headquartered in Ireland under EU law. The DPC’s intervention comes amid increased scrutiny of AI development practices across the EU by major tech companies. Recently, the DPC sought an order to halt or suspend X’s data processing activities related to AI system development, training, and refinement. This situation illustrates the growing tension across nearly all EU states between AI advancements and ongoing data protection concerns. However, it appears that regulators and the court issued the order too late. In response to the lawsuit, X, owned by Elon Musk, reported that Grok—an AI chatbot—allowed users to opt out of having their public posts used. Judge Leonie Reynolds noted that X began processing European users’ data for AI training on May 7, but the opt-out option was not introduced until July 16, and it wasn’t immediately available to all users. Consequently, there was a period during which data was used without user consent. X’s legal team assured the court that data obtained from EU users between May 7 and August 1 will not be used while the DPC’s order is under review. X is expected to file opposition papers against the suspension order by September 4, potentially sparking a court battle that could have significant repercussions across the EU. X has not remained silent on the matter. In a statement from the company’s Global Government Affairs account on X, the DPC’s order was described as “unwarranted, overbroad, and unfairly targeting X without any justification.” The company also expressed concerns that the order could hinder efforts to maintain platform safety and limit its use of technologies in the EU, underscoring the complex balance between regulatory compliance and operational viability that tech companies must navigate in today’s digital landscape. X emphasized its proactive approach in working with regulators, including the DPC, regarding Grok since late 2023. The company claims to have been fully transparent about using public data for AI models, providing necessary legal assessments, and engaging in extensive discussions with regulators. This regulatory action against X is not an isolated incident. Other tech giants have faced similar scrutiny in recent months. Meta Platforms recently postponed the launch of its Meta AI models in Europe following advice from the Irish DPC. Similarly, Google agreed to delay and modify its Gemini AI chatbot earlier this year after consultations with the Irish regulator. These events signal a shift in the regulatory landscape concerning AI and data usage in the EU. Regulators are taking a more active role in overseeing how tech companies use user data for AI training and development, reflecting growing concerns about data privacy and the ethical implications of AI advancement. As legal proceedings continue, the outcome of this case could set important precedents for AI development regulation in the EU, potentially influencing global data protection standards in the AI era. Both the tech industry and privacy advocates will be closely monitoring this situation, recognizing its potential to shape the future of AI innovation and data privacy regulations.

Anthropic has introduced Claude 3.5 Sonnet, a mid-tier model that not only outperforms its competitors but also surpasses Anthropic’s current top-tier model, Claude 3 Opus, in various evaluations. Claude 3.5 Sonnet is now freely accessible on Claude.ai and the Claude iOS app, with higher rate limits for subscribers of Claude Pro and Team plans. It’s also available through the Anthropic API, Amazon Bedrock, and Google Cloud’s Vertex AI, priced at $3 per million input tokens and $15 per million output tokens, with a 200K token context window. Anthropic asserts that Claude 3.5 Sonnet “sets new industry benchmarks for graduate-level reasoning (GPQA), undergraduate-level knowledge (MMLU), and coding proficiency (HumanEval).” The model exhibits enhanced capabilities in understanding nuance, humor, and complex instructions, while excelling at producing high-quality content with a natural tone. Operating at twice the speed of Claude 3 Opus, Claude 3.5 Sonnet is well-suited for complex tasks such as context-sensitive customer support and multi-step workflow orchestration. In an internal agentic coding evaluation, it solved 64% of problems, significantly outperforming Claude 3 Opus, which solved 38%. Additionally, the model demonstrates improved vision capabilities, outperforming Claude 3 Opus on standard vision benchmarks. This improvement is particularly evident in tasks requiring visual reasoning, such as interpreting charts and graphs. Claude 3.5 Sonnet can accurately transcribe text from imperfect images, a valuable feature for industries like retail, logistics, and financial services.

To avoid major vulnerabilities that could lead to sensitive information theft, financial fraud, or operational disruption, it’s crucial to have visibility into your custom application logic and data flows. While perimeter firewall defenses are still necessary, relying solely on them to protect increasingly powerful web properties is a risky gamble with serious consequences. Incorporating specialized web application firewalls (WAFs) that are designed to analyze requests within the full context of your application environments – and enhanced with AI for greater accuracy – allows you to secure your systems and confidently develop advanced digital capabilities. By adopting a layered defense-in-depth strategy that combines both network and application-level protections, you can safely deliver the seamless, personalized digital experiences that are key to building lasting customer relationships and achieving operational excellence in 2024. If you run any online services, chances are you already have traditional firewall protection guarding your overall network. These firewalls filter incoming traffic based on a set of predefined rules that primarily focus on protocol, port number, IP address ranges, and basic connection state. Traditional firewalls provide effective perimeter protection against classic network-focused cyberthreats. However, they lack the context needed to understand the application logic, user workflows, and data structures unique to custom web apps and APIs. These firewalls merely scan network packets as they arrive and decide whether to allow or block them, leaving them susceptible to the evolving tactics of AI-powered attackers.

Google has introduced its new Pixel 9 smartphones, highlighting their advanced AI features. The release of these devices comes earlier than usual, as Google typically unveils new Pixel models in the autumn. The changes in these new models are more revolutionary compared to previous versions. These smartphones boast enhanced integrations of Google’s AI technology. One notable feature allows users to search for information and images within their screenshots, thanks to deeper integration. Additionally, some features are accessible as overlays from other apps through the Gemini chatbot. The Pixel 9 series includes several models. The base model, Pixel 9, features a 6.3-inch screen and is priced at $799. The Pixel 9 Pro XL, a larger option, comes with a 6.8-inch screen. The Pixel 9 Pro offers an enhanced camera system at a higher price, while the Pixel 9 Pro Fold is a foldable variant. Google announced at the event that the Pixel 9 and Pixel 9 Pro XL will ship in late August. The Pixel 9 Pro and Pixel 9 Pro Fold are scheduled for a September release, with preorders starting August 13. The event featured a live demo showcasing Gemini’s new conversation features. Google also highlighted updates to the phone’s exterior design, the advanced camera system, and the integration of the new Tensor G4 chip. Additionally, Google unveiled the Pixel Watch 3 smartwatch and Pixel Buds Pro 2 wireless earbuds. The smartwatch includes a heart rate tracking feature that calls emergency services if the heart rate stops, available in the UK and the EU.

Primate Labs has officially released Geekbench AI, a benchmarking tool tailored for machine learning and AI-focused workloads. Geekbench AI 1.0 represents the culmination of years of development and collaboration with customers, partners, and the AI engineering community. Formerly known as Geekbench ML during its preview phase, the benchmark has been rebranded to match industry terminology and clarify its purpose. Now available on the Primate Labs website for Windows, macOS, and Linux, Geekbench AI can also be accessed via the Google Play Store and Apple App Store for mobile devices. This new benchmarking tool aims to provide a standardized method for assessing and comparing AI capabilities across various platforms and architectures. It features a distinctive approach by offering three overall scores, capturing the complexity and diversity of AI workloads. Geekbench AI’s three-score system addresses varied precision levels and hardware optimizations in modern AI implementations. This multi-dimensional approach provides deeper insights into AI performance across different scenarios for developers, hardware vendors, and enthusiasts. A key feature of Geekbench AI is the addition of accuracy measurements for each test, reflecting that AI performance involves both speed and result quality. By combining speed and accuracy metrics, the benchmark offers a comprehensive view of AI capabilities, highlighting the trade-offs between performance and precision. Geekbench AI 1.0 supports a wide range of AI frameworks, including OpenVINO on Linux and Windows, and vendor-specific TensorFlow Lite delegates such as Samsung ENN, ArmNN, and Qualcomm QNN on Android. This extensive framework support ensures the benchmark remains relevant with the latest tools and methodologies. The benchmark uses diverse and extensive datasets, enhancing accuracy evaluations and better representing real-world AI use cases. Each workload in Geekbench AI 1.0 runs for at least one second, enabling devices to reach peak performance while reflecting real-world application bursts. Primate Labs has provided detailed technical descriptions of the workloads and models used in Geekbench AI 1.0, demonstrating their commitment to transparency and industry-standard testing. Integrated with the Geekbench Browser, the benchmark allows for easy cross-platform comparisons and result sharing. Primate Labs plans regular updates to Geekbench AI to adapt to market changes and emerging AI features. The company believes the benchmark’s current reliability makes it suitable for professional workflows, with major tech companies like Samsung and Nvidia already using it.

OpenAI has introduced fine-tuning capabilities for its GPT-4o model, a much-anticipated feature for developers. To make the offer more enticing, OpenAI is offering one million free training tokens daily per organization until 23rd September. Fine-tuning GPT-4o with custom datasets can enhance performance and lower costs for specific applications. This process allows for precise control over the model’s responses, enabling customization of structure, tone, and adherence to detailed, domain-specific instructions. Developers can achieve significant results with training datasets as small as a few dozen examples, making this feature accessible for improvements in various fields, from complex coding to nuanced creative writing. GPT-4o fine-tuning is now available to all developers across all paid usage tiers. Training is priced at $25 per million tokens, with inference costs set at $3.75 per million input tokens and $15 per million output tokens. Additionally, OpenAI is offering GPT-4o mini fine-tuning with two million free daily training tokens until 23rd September. To access this, developers can select “gpt-4o-mini-2024-07-18” from the base model dropdown on the fine-tuning dashboard.