OpenAI's Mystery Model Stuns Everyone

GPT2-Chatbot is incredible and was quickly removed

An unexplained AI model, dubbed "gpt2-chatbot," has surfaced on the LMSYS Chatbot Arena, sparking intrigue for its impressive abilities that rival or perhaps surpass OpenAI's current top model, GPT-4. However, experts suggest it isn't advanced enough to be the anticipated GPT-5, leading to conjecture it could be an interim update like a GPT-4.5. OpenAI CEO Sam Altman's recent cryptic social media activity hints at a connection to this new AI.

Sponsor

Vultr is empowering the next generation of generative AI startups with access to the latest NVIDIA GPUs.

Try it yourself when you visit getvultr.com/forwardfutureai and use promo code "BERMAN300" for $300 off your first 30 days.

• Reid Hoffman interviewed his AI deepfake, which was pretty convincing - Reid Hoffman, LinkedIn co-founder and an AI enthusiast, conducted an unusual experiment by interviewing a digital replica of himself, dubbed Reid AI. The AI, developed using OpenAI's GPT-4 and trained with Hoffman's extensive public domain work, engaged in a conversation with him that was simultaneously bizarre and intriguing. They touched upon various topics including AI regulation, its potential, and personal improvements to Hoffman’s LinkedIn profile. Hoffman acknowledged the necessity of ethical guidelines for deepfakes, comparing AI's occupational disruption to the historical impact of the steam engine. Notably, Reid AI even advocated for governmental regulation to ensure fair and beneficial AI deployment, aligning with Hoffman's support for using AI to enhance human progress and his belief in "blitzscaling" AI development for rapid growth.

  

• CEOs of OpenAI, Google and Microsoft to join other tech leaders on federal AI safety panel - The U.S. Department of Homeland Security is forming a panel of CEOs from top tech and industry companies to advise on utilizing artificial intelligence to protect critical infrastructure, including against AI-powered threats. Notable members include leaders from Google, Microsoft, and OpenAI, among other tech giants, defense firms, and civil rights organizations. The AI Safety and Security Board, established under a 2023 executive order by President Biden, will guide sectors such as telecommunications and utilities on responsible AI use and preparedness for AI-related disruptions. This action highlights the government's integrated approach in absence of a specific national AI law. This board follows national concerns about the potential use of AI, including deepfakes, in election interference.

• Meet the AI Expert Advising the White House, JP Morgan, Google and the Rest of Corporate America - Ethan Mollick, a business school professor at the University of Pennsylvania, has emerged as a prominent AI expert advising entities such as the White House, JPMorgan, and Google. His deep involvement with AI began with the release of ChatGPT in November 2022, which he quickly embraced and explored. Mollick actively shared his findings and experiences on social media and incorporated AI into his teaching at Wharton, where he observed its impact on his MBA students. His enthusiastic adoption and thorough understanding of large-language models have made him a key resource for both policymakers and corporate leaders.

• Apple intensifies talks with OpenAI for iPhone generative AI features - Apple is intensifying talks with OpenAI to utilize the startup's technology in new iPhone features, specifically focusing on integrating OpenAI features into the iOS 18 update. Negotiations also involve discussions with Google for licensing its Gemini chatbot technology for potential integration into Apple's products. While Apple has not finalized any agreements yet, it is considering partnerships with both OpenAI and Google, or potentially another provider, to enhance its AI capabilities. These developments precede Apple's Worldwide Developers Conference, where the company plans to unveil new AI software and services emphasizing seamless integration and privacy protections.

  

• TSMC unveils 1.6nm process technology with backside power delivery, rivals Intel's competing design - TSMC introduced its 1.6nm-class A16 semiconductor manufacturing process, boasting significant performance enhancements over its N2P predecessor. Emphasizing efficiency and performance, the 1.6nm process incorporates gate-all-around (GAA) nanosheet transistors alongside a breakthrough: the backside power delivery network (BSPDN), known as Super Power Rail (SPR). This advance is tailored for AI and HPC processors, aiming to improve transistor density by 7%-10% and reduce power consumption by 15%-20% while increasing clock speeds by up to 10%. Volume production is slated for the second half of 2026 with market introduction possibly in 2027, setting up competition with Intel's 14A node.

• Most global tech leaders see their companies unprepared for AI - Artificial intelligence (AI) has emerged as a primary focus for IT leaders globally, yet businesses struggle with preparedness for its effective deployment, as highlighted in a Lenovo Group Ltd. survey of 750 chief information officers (CIOs). These CIOs, representing diverse global markets, have identified significant gaps in computing infrastructure and corporate policies regarding ethical AI usage, and 42% do not foresee a return on their AI investments for at least two years. Despite these challenges, there is an impending increase in AI investments, with 96% of CIOs expecting to boost their spending over the next year, although demonstrating tangible benefits from these investments remains difficult. Lenovo's survey underscores the intensifying pressure on IT leaders to not only adopt AI but also to validate the financial viability of these technologies in their operations.

• Elon Musk’s xAI startup closes in on $6 Billion Fundraising - Elon Musk's AI startup, X.AI Corp., is close to securing $6 billion in a funding round that would elevate the company's valuation to $18 billion. Sequoia Capital, among other investors, is participating in this financing effort. This funding move is part of xAI's strategy to challenge major players in the AI industry, including OpenAI, of which Musk was a founding member. The fundraising aims not only to fuel competition but also to navigate Musk's increasingly complex relationship with OpenAI.

• Coca-Cola is betting AI can help it compete against Pepsi and Red Bull - Coca-Cola is investing in AI to stay competitive in the beverage market, particularly against rivals Pepsi and Red Bull. The company is using machine learning to analyze consumer behavior, optimize production, and improve customer service. Additionally, Coca-Cola is leveraging AI to enhance its e-commerce capabilities and improve its relationships with retailers. By embracing AI, Coca-Cola aims to stay ahead in the rapidly changing retail landscape and maintain its market lead.

  

Sponsor

Pinecone’s serverless vector database helps you deliver remarkable GenAI applications faster at up to 50x lower cost. Learn more about Pinecone serverless here.

Awesome Research Papers

Make Your LLM Fully Utilize the Context - This abstract introduces FILM-7B, a large language model designed to address the "lost-in-the-middle" challenge where models lose track of vital information in long text segments. It proposes an "information-intensive" training method using a synthesized dataset to enhance the model's accuracy in extracting relevant data from lengthy contexts, ranging from 4K to 32K tokens. FILM-7B’s effectiveness is validated through specialized probing tasks and its performance is shown to improve on long-context benchmarks while retaining quality on short-context tasks. The work, including FILM-7B, is available on GitHub for further exploration.

Many-Shot In-Context Learning - The paper explores the extension of in-context learning (ICL) with large language models (LLMs) to include hundreds or thousands of examples, termed the many-shot regime. It demonstrates that many-shot ICL can significantly improve performance across various tasks, but is limited by the availability of human-generated examples. To address this, the authors introduce Reinforced and Unsupervised ICL, which use model-generated rationales and domain-specific questions without rationales, respectively, showing effectiveness especially in complex reasoning tasks. The research also highlights the potential of many-shot learning to override pretraining biases and the limitations of next-token prediction loss as a performance indicator.

Chinese team unveils first text-to-video AI model on par with Sora - At the 2024 Zhongguancun Forum, a Chinese innovation in AI was unveiled: Vidu, a text-to-video model created by ShengShu-AI and Tsinghua University, matching the capabilities of the earlier-released OpenAI's Sora. Vidu uses a proprietary Universal Vision Transformer architecture to produce 1080P video clips from text prompts. This system came two months after Sora and boasts the ability to create realistically detailed scenes that adhere to physical laws. Its design, which integrates Diffusion and Transformer AI models, prioritizes understanding Chinese cultural elements and can generate distinctly Chinese imagery.