U.S. Export Rules Prompt Nvidia to Create China-Specific AI Processor

Nvidia Preparing Version of New Flagship AI Chip for Chinese Market

Nvidia is developing a version of its new flagship AI chip, the B20, for the Chinese market, compliant with U.S. export controls. This follows the unveiling of its Blackwell chip series in March, with the B200 offering significantly enhanced performance. Nvidia will collaborate with Inspur for the distribution, with shipments expected by Q2 2025. The initiative aims to counteract the impact of U.S. sanctions and strengthen Nvidia's market presence in China, where revenue has recently declined. Despite tighter export controls, Nvidia's tailored chips for China have seen increasing sales, with high expectations for continued market growth.

Sponsor

The HP Elitebook 1040 G11 powered by Intel Core Ultra Processor that unlocks AI experiences. Innovative design and real power for real work. Check out the HP Elitebook 1040 G11 powered by Intel Core Ultra Processor today! https://bit.ly/4b6FJQT

• OpenAI Holds Talks with Broadcom About Developing New AI Chip - OpenAI is in discussions with Broadcom and other chip designers to develop a new AI chip, as reported by the Information. This initiative aims to address the shortage of expensive GPUs necessary for creating AI models like ChatGPT and DALL-E3. The Microsoft-backed company is hiring former Google employees who developed the tensor processing unit and is planning to create an AI server chip. Additionally, CEO Sam Altman has plans to raise billions to establish semiconductor manufacturing factories with potential partners including Intel, Taiwan Semiconductor Manufacturing Co, and Samsung Electronics.

• The Data That Powers A.I. Is Disappearing Fast - An MIT-led study by the Data Provenance Initiative observed that major web sources for training artificial intelligence are starting to limit their data's availability. Analyzing 14,000 web domains from A.I. training sets C4, RefinedWeb, and Dolma, the research identified an "emerging crisis in consent." Among these domains, 5% have enacted restrictions, which rise to 25% for higher-quality sources, mainly via the Robots Exclusion Protocol and terms of service prohibitions. This trend marks a significant shift, creating challenges for A.I. development across commercial, research, and academic spheres.

• It May Soon Be Legal to Jailbreak AI to Expose How it Works - The proposed exemption to the Digital Millennium Copyright Act (DMCA) aims to empower researchers by allowing them to bypass the terms of service on artificial intelligence tools for investigative purposes. This exemption would enable them to scrutinize AI systems for biases, analyze training data, and identify outputs that might be harmful, thus facilitating a better understanding of these complex systems and holding them accountable for algorithmic fairness.

• How AI Brought 11,000 College Football Players to Digital Life in Three Months - Electronic Arts (EA) used advanced AI technology to create digital replicas of 11,000 college football players for "EA Sports College Football 25," overcoming the logistical challenges of traditional 3D scanning. This approach allowed EA to generate 3D avatars quickly by using photos provided by schools, marking a significant advancement from the generic player models of past games. The AI process involved iterative improvements by artists, ensuring accurate player likenesses. EA's innovative techniques also included detailed recreations of 134 college stadiums and have laid the groundwork for future sports titles and annual game updates.

  

• AI terminology, explained for humans - The article offers a comprehensive guide to artificial intelligence (AI) terminology, providing clear explanations for common AI jargon. AI, the discipline focused on creating systems that think like humans, is now a prevalent concept in technology, often touted as a key selling point by numerous companies. The field includes specific areas like machine learning, where systems are trained on data to make predictions, and artificial general intelligence (AGI), denoting AI that matches or surpasses human intelligence. Other terms explained include generative AI, responsible for creating new content, and "hallucinations," where AI generates confident, incorrect responses based on flawed training data.

• OpenAI’s latest model will block the ‘ignore all previous instructions’ loophole - OpenAI researchers have developed a new technique called "instruction hierarchy" to counter users' attempts to manipulate AI with conflicting directions, like the 'ignore all previous instructions' meme. This method prioritizes developer-set instructions over user inputs, improving the AI's resistance to misuse. OpenAI's budget-friendly GPT-4o Mini is the first model to employ this safety feature. The technique is integral to OpenAI's goal of creating automated agents capable of managing digital aspects of users' lives safely. Amidst internal and external safety and transparency concerns at OpenAI, this new mechanism is a step towards rebuilding trust and enhancing control over AI behavior, distinguishing between system-level commands and potentially harmful user prompts.

• Illicit Chip Flows to Russia Slow, but China and Hong Kong Remain Key Transshipment Hubs - The flow of restricted goods, including semiconductors, from China and Hong Kong to Russia has decreased by 20% in early 2024, as revealed by undisclosed U.S. Commerce Department data. Despite this decline, Hong Kong remains a central hub for sanctions evasion, with numerous companies facilitating the transfer of high-end chips and other advanced components to support Russia's military. Efforts by the U.S. and its allies to curb this trade include aggressive enforcement and engagement with manufacturers. However, Hong Kong and Chinese authorities have not fully aligned with U.S. unilateral sanctions, maintaining their stance of only enforcing United Nations Security Council mandates.

  

• SAG-AFTRA can now call immediate strike if game companies refuse AI protections - The SAG-AFTRA National Board has enabled Duncan Crabtree-Ireland to call a strike if required to protect voice actors under the Interactive Media Agreement. As the union representing around 160,000 media professionals, SAG-AFTRA is advocating for better AI protections, safety measures, and wages in line with inflation during renegotiations. With significant opposition from employers on AI-related clauses, negotiations are at a standstill, prompting the authorization for immediate strike action. The union's members have already shown overwhelming support for a strike, with a 98% affirmative vote last year. If an agreement isn't reached, SAG-AFTRA is prepared to lead its members to the picket lines.

• UTA Signs AI-Focused Native Foreign Creative Agency & Production Company - UTA has signed the creative agency and production company, Native Foreign, known for Emmy nominations and being a WBENC-certified women-owned business. Specializing in AI-driven creative storytelling, Native Foreign recently garnered attention for their Toys“R”Us brand film, which rapidly progressed from concept to completion thanks to AI technology. Their CCO, Nik Kleverov, an Emmy nominee for "Narcos," is recognized for pioneering AI applications in creative work. Native Foreign boasts a portfolio including Netflix, Amazon Prime, and Showtime collaborations, and has earned Emmy nods for the literacy initiative, Storyline Online. UTA aims to enhance Native Foreign's integration of AI into film, entertainment, and branded content.

• AI Startup Cohere Valued at $5.5 Billion in New Funding Round - Cohere Inc., a Canadian AI startup, has reached a valuation of $5.5 billion after raising $500 million in a Series D funding round. Unlike its competitors, Cohere focuses on developing large language models for business applications rather than consumer-facing AI tools. The company’s technology is used by various clients, including Notion Labs and Oracle, to enhance their products with generative AI. Founded in 2019 and led by AI expert Aidan Gomez, Cohere has rapidly grown, doubling its valuation and headcount within a year, while maintaining its base in Toronto.

• Meta is working on AI video generation for Imagine feature - Meta is advancing its AI capabilities by introducing video generation functions to its Meta AI platform. Users can engage with text-based prompts to create videos, potentially for use on Instagram Reils, reflecting a storytelling approach whereby individual scenes can be produced as distinct video clips. Meanwhile, there are indications of an upcoming premium model, LLama 3 405B, which might be part of Meta AI's rumored premium service. Additionally, references to an AI Studio point towards a developer-focused tool, possibly akin to platforms like OpenAI Playground or Google AI Studio, aimed at integrating Meta's AI models into external applications.

  

Awesome Research Papers

• Design and Control of a Bipedal Robotic Character - This work focuses on the intersection of legged robotics and entertainment, introducing a bipedal robot designed for expressive and artistic movement. A reinforcement learning-based control architecture enables the robot to perform dynamic locomotion, while an animation engine generates real-time gestures for engaging performances. Through an operator interface, these movements can be controlled live, allowing for interactive shows. The research promises to enhance human-robot interaction and finds potential applications beyond entertainment.

  

• As Generative Models Improve, People Adapt Their Prompts - In an online experiment with 1,891 participants and over 18,000 prompts, researchers explored how prompting evolves with advanced AI models. Participants used DALL-E 2, DALL-E 3, or DALL-E 3 with automatic prompt revision to replicate target images. Results showed that DALL-E 3 users wrote longer and more descriptive prompts, leading to higher performance than DALL-E 2. However, the use of automated prompt revisions reduced these benefits by 58%. The study concludes that as AI models improve, the practice of crafting effective prompts remains crucial.

• BRIGHT: A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval - BRIGHT is the first text retrieval benchmark designed to tackle complex queries that demand in-depth reasoning beyond simple keyword matching. It comprises 1,398 real-world queries from various domains, requiring the understanding of logic and syntax in areas like software engineering and earth sciences. State-of-the-art models struggle on BRIGHT, with the best model scoring only 18.0 in normalized Discounted Cumulative Gain (nDCG@10), compared to 59.0 on other benchmarks. Increasing model performance by up to 12.2 points is possible by integrating queries with reasoning from large language models. BRIGHT also withstands data leakage issues, maintaining consistent performance despite potential training data overlap.

• Benchmarking Trustworthiness of Multimodal Large Language Models: A Comprehensive Study - The paper outlines the establishment of MultiTrust, a comprehensive benchmark designed to evaluate the trustworthiness of Multimodal Large Language Models (MLLMs). It addresses five primary aspects: truthfulness, safety, robustness, fairness, and privacy, through 32 tasks and self-curated datasets. Testing 21 modern MLLMs, it unveils various trust issues, such as vulnerability to visual confusions, multimodal jailbreaking, adversarial attacks, privacy breaches, and biases. These findings suggest that multimodality can exacerbate risks inherent to base language models. The study also introduces a scalable toolbox to assist in trustworthy MLLM research, with resources available online.

• Scaling Retrieval-Based Language Models with a Trillion-Token Datastore - This paper presents an exploration into scaling laws of language models (LMs) concerning inference data size. It demonstrates that larger datastores improve LMs both in language tasks and various applications, sometimes enabling smaller models to outperform larger ones. The study introduces MassiveDS, a record-sized 1.4 trillion-token dataset, used to evaluate the benefits of expanding datastores. Additionally, the paper investigates the impact of retrieval improvements, quality filtering, and other factors. The findings underscore the importance of datastore size in optimizing LM efficiency and performance, and the research materials have been made available for public use.

Apple shows off open AI prowess: new models outperform Mistral and Hugging Face offerings - Apple's research team, collaborating with academic and industry partners, announced the release of the open-source DataComp for Language Models (DCLM) project on Hugging Face, launching two transformer-based English language models. The first, a 7 billion parameter model, demonstrates a 63.7% 5-shot accuracy on MMLU, outperforming the prior state-of-the-art and boasting a significant reduction in compute required for training. Additionally, a smaller 1.4 billion parameter model shows superior results when compared to its peers. Both models, which herald improvements in data curation techniques, come with open-source credentials including model weights, training code, and pretraining datasets, with the smaller model released under the more permissive Apache 2.0 license. Notable is the emphasis on the potential biases and non-readiness of these early research models for integration into Apple products.

Introducing Llama-3-Groq-Tool-Use Models - Groq is Fast AI Inference - Groq and Glaive have released two new open-source models, Llama-3-Groq-70B-Tool-Use and Llama-3-Groq-8B-Tool-Use, designed for advanced tool use and function calling. The Llama-3-Groq-70B-Tool-Use model leads the Berkeley Function Calling Leaderboard with 90.76% accuracy, while the 8B variant ranks third with 89.06%. These models, trained with ethical data and advanced optimization techniques, are available on GroqCloud and Hugging Face. To maximize AI system performance, a hybrid approach combining these specialized models with general-purpose language models is recommended. The models aim to enhance AI applications across various domains and encourage community innovation.

Introducing the Coalition for Secure AI, an OASIS Open Project - The Coalition for Secure AI (CoSAI) was unveiled at the Aspen Security Forum, sponsored by the OASIS global standards organization. It unites a diverse group of stakeholders to standardize AI security practices. Top industry firms like Google, IBM, and Microsoft are among its founding sponsors. CoSAI will focus on creating Secure-by-Design AI systems, mitigating risks like model theft and data poisoning, and fostering secure AI application deployment. CoSAI is governed by a board and a technical committee, aimed at addressing the current disjointed state of AI security. They plan to establish workstreams on software supply chain security for AI, adapting cybersecurity for emerging AI, and AI security governance. CoSAI invites open-source community contributions and seeks more sponsors to join the initiative.

EverArt - EverArt is an all-encompassing platform designed for the creation and management of AI-generated art, especially for brand management. It features a personalized workspace for project management and collaboration, a variety of AI models for generating diverse art styles, and organizational tools like boards to visualize multiple projects. Additionally, EverArt includes a billing system to manage subscriptions and payments, making it a robust solution for artists and designers working with AI technology.

Google developing Prompts Gallery for user-generated prompts - Google is developing a Prompts Gallery, poised to enable user submissions and inspiration sharing for prompts. Features spotted include a "Get Inspired" button for access to a curated collection and a "Submit to Gallery" option funneling submissions via a Google Form. While current access is limited, there's speculation about a public release. This initiative complements Gemini, Google's project committed to boosting artificial intelligence prowess, focusing on natural language understanding and integration with existing services, signaling continuous AI advancement efforts.

Perplexity adds easy access to site specific searches.