OpenAI Releases “Security” Proposal, Which Is Extremely Concerning

GPU Tracking, Inference Authorization, and more…

• OpenAI Says It Can Now Detect Images Spawned by Its Software—Most of the Time - OpenAI has introduced a novel detection tool designed to identify images generated by its advanced text-to-image generator, DALL-E 3. The tool boasts impressive accuracy in detecting DALL-E 3-generated images, but may struggle with subtle alterations to an image. This development marks a significant step in AI self-awareness, as the ability to distinguish AI-generated content from human-created material becomes increasingly crucial. The tool's implementation is a milestone in AI development, reflecting the ongoing pursuit of AI companies to enhance their technology's ability to detect its digital footprint and maintain the integrity of digital content.

  

• Apple Reportedly Building M2 Ultra and M4-Powered AI Servers - Apple is slated to unveil iOS 18 at its WWDC keynote on June 10, 2024. Anticipation builds as rumors suggest enhancements across multiple apps including Apple Music, Maps, Calculator, Messages, Notes, and Safari. These rumored updates have been reported by multiple sources, with MacRumors and Bloomberg's Mark Gurman leading the discourse. Specific improvements are expected for Apple Maps, with at least two new features on the horizon.

• Secret Apple 'Project ACDC' to pioneer AI chips for data centers - Apple is secretly developing "Project ACDC," an initiative to create proprietary AI chips for data center applications. This project aims to optimize AI applications within Apple's data centers for future versions of its platforms, leveraging the company's expertise in custom chip development for devices like iPhones and Mac computers. The move is seen as a response to the intensifying AI competition, particularly from Microsoft and Meta, which have rapidly adopted generative AI following the success of OpenAI's ChatGPT. Apple plans to invest over $5 billion in AI server enhancements over the next two years, aiming to match the technological capabilities of industry leaders. The company has reportedly secured the entire production capacity of TSMC's 3-nanometer processors, which were initially thought to be for iPhone and Mac chips, but are now believed to be for a specialized AI processor with mass production expected to start in the latter half of 2025.

  

• Our Approach to Data and AI - OpenAI's approach to data and AI emphasizes the importance of respecting the choices of creators and content owners, ensuring that AI systems benefit and respect their preferences. To achieve this, OpenAI is developing Media Manager, a tool that will enable creators and content owners to manage how their works are used in AI, aiming to set a standard across the AI industry by 2025. The company's AI models are designed to learn from relationships in information, not store data like a database, and are trained on broad and diverse datasets to benefit as many people as possible. OpenAI collaborates with creators, content owners, and regulators to develop products that benefit users, creators, and publishers, and to enhance the user experience.

• Stanford AI leader Fei-Fei Li building 'spatial intelligence' startup - Fei-Fei Li, a prominent AI leader from Stanford, is building a startup focused on "spatial intelligence," a technology that enables machines to understand and interact with their physical environment. The startup aims to develop AI systems that can perceive and navigate spaces, similar to how humans do. This technology has the potential to revolutionize various industries, from robotics and autonomous vehicles to smart homes and cities.

• AI Copilots Are Changing How Coding Is Taught - Generative AI is reshaping the landscape of software development, influencing both industry practices and academic curricula. Programmers are utilizing AI-driven coding aids to streamline their work, and AI-focused jobs are on the rise. Academically, computer science students are turning to generative AI to grasp intricate concepts and enhance learning, spurring educators to adapt teaching methodologies. Fundamental skills in education are shifting, with an escalated focus on problem-decomposition and software testing. While educators are integrating such AI tools to enrich higher-order thinking and software design instruction, they remain wary of students becoming overly reliant on AI, with concerns about AI's potential for generating biased or incorrect code. The academic community is striving to balance AI technology integration with critical learning, aiming to bridge the gap between university education and industry requirements.

  

• Microsoft and Brookfield Sign Biggest-Ever Clean Power Deal - Microsoft has entered into a significant green power agreement with Brookfield Renewable. This deal marks a major step towards Microsoft's goal of becoming carbon negative by 2030. The partnership will enable Microsoft to power its data centers with renewable energy, reducing its carbon footprint and contributing to a more sustainable future.

• Meet the Woman Who Showed President Biden ChatGPT—and Helped Set the Course for AI - Arati Prabhakar, the director of the White House Office of Science and Technology Policy (OSTP) and President Biden's chief science and technology adviser, has been instrumental in shaping the administration's policies on AI, semiconductors, and cancer research. In March 2023, she demonstrated the capabilities of ChatGPT to President Biden, which inspired a comprehensive executive order on AI regulation six months later.

  

• Microsoft Develops Internet-Free AI Model for Spy Agencies - Microsoft has developed a generative AI model specifically tailored for U.S. intelligence agencies, capable of operating in a completely isolated environment without internet connectivity. This is a first for Microsoft and is intended to serve the secure needs of government entities like the CIA. The AI, distinguishable from cloud-dependent models, can analyze highly sensitive data on a secure network. Microsoft's CTO for strategic missions, William Chappell, confirmed this, aligning with intelligence officials' desire to employ advanced AI tools for security purposes. The model, known as MAI-1, boasts an impressive 500 billion parameters and competes with other major industry players.

• Meta Will Let Advertisers Create Campaigns Using New Generative AI Tools - Meta Platforms Inc., the parent company of Facebook and Instagram, is introducing new tools that enable advertisers to create marketing materials using generative AI prompts. These tools can generate new images of a product from a single uploaded photo, and will soon allow marketers to create different versions of an image using text prompts. Additionally, Meta's AI can generate text for advertisements, including headlines and primary ad copy, with the goal of reflecting a brand's voice and tone. This expansion of AI capabilities is part of CEO Mark Zuckerberg's broader push into AI technology, aiming to make marketing more accessible and cost-effective for small- and medium-sized businesses.

• Microsoft's Copilot AI is stealing one of Midjourney's best features - Microsoft is integrating new features into its Copilot product, including a rewrite function that enhances prompts to produce more satisfactory results, similar to techniques used by pioneering AI art platform Midjourney. Additionally, Microsoft introduces a sharing capability for optimized prompts and a Catch Up feature for recommending actionable steps. The rewrite tool aims to assist users in crafting effective prompts by adding rich context, essentially turning users into "prompt engineers." Details on the execution of these features remain vague, but they are expected to be guided by user data and context. These additions to Copilot are scheduled to launch in the near future.

  

• Biden Will Announce Microsoft’s $3.3 Billion AI Datacenter In Wisconsin—On Same Site As Trump’s Failed Foxconn Factory - On May 8, 2024, President Joe Biden announced Microsoft's plan to invest $3.3 billion in a new AI datacenter in Racine, Wisconsin. This initiative is expected to generate 2,300 construction jobs and 2,000 permanent positions, with commitments to train 2,000 individuals in datacenter and AI adoption. Aimed at reinforcing U.S. manufacturing in various high-tech sectors, the project contrasts with a failed $10 billion project by Foxconn in the same location, heralded by the previous administration but which drastically fell short of job creation promises.

• OpenAI says it's building a tool to let content creators 'opt out' of AI training - OpenAI is creating Media Manager, a tool designed to help content creators control how their works are used in training generative AI models, with plans to launch it by 2025. This development aims to address criticism and legal challenges regarding OpenAI's data sourcing practices, offering creators a way to specify their preferences for inclusion or exclusion in training datasets. Despite these efforts, some creators believe that OpenAI’s measures are insufficient, finding the opt-out process cumbersome and incomplete.

• Who stole the show at this year’s Met Gala? The AI Deepfakes - The Met Gala, a premier fashion event, has become a prime target for deepfakes. This year's event saw a rise in manipulated imagery, categorized into three groups: amusing forgeries, imaginative fashion designs, and deceptive deepfakes. The latter, unlabeled and created to sow confusion, led to instances like Katy Perry's mother being fooled into believing an AI-generated image of her daughter attended the event. The increasing accessibility of visual AI generators like DALL-E 3 has contributed to this trend.

  

• AI at Work Is Here. Now Comes the Hard Part - The 2024 data indicates a significant increase in generative AI use at work, with 75% of global knowledge workers now employing it. While AI's potential for improving bottom-line growth is acknowledged, many organizations lack concrete plans for large-scale implementation, leading to leadership paralysis despite acknowledging the inevitability of AI's role in business. AI adoption parallels historical tech disruptions, suggesting transformative impact for businesses that leverage AI effectively. The labor market is poised to evolve, with AI skills potentially becoming more valuable than experience and amplifying career opportunities for many workers. Microsoft and LinkedIn's comprehensive study of 31,000 people across 31 countries, along with an analysis of Microsoft 365 productivity data and LinkedIn hiring trends, offers leaders critical insights and recommended actions regarding AI's influence on the future of work.

Awesome Research Papers

Hallucination of Multimodal Large Language Models: A Survey - This comprehensive survey delves into the issue of hallucinations in multimodal large language models (MLLMs), where generated outputs may not align with visual content. As hallucination undermines the practical use and reliability of MLLMs, significant efforts are underway to detect, evaluate, and correct these errors. The survey provides an in-depth analysis of the causes of hallucinations, the benchmarks for evaluating them, and the methods for mitigation, aiming to foster a deeper understanding and to guide future research in improving MLLMs' robustness and reliability.

xLSTM: Extended Long Short-Term Memory - The 1990s introduced Long Short-Term Memory (LSTM) models which were fundamental to the development of Large Language Models (LLMs). Transitioning to an era dominated by Transformers, researchers propose a question about the potential of scaling LSTMs. The study revisits and enhances LSTM by integrating exponential gating and modifying its memory structure to include scalar (sLSTM) and matrix (mLSTM) memory types. By doing so, it creates xLSTM blocks used in scalable architectures that show competitive performance against contemporary Transformers and State Space Models in language modeling tasks, demonstrating that advanced LSTMs can still be relevant and effective.

Agent Hospital: A Simulacrum of Hospital with Evolvable Medical Agents - This paper presents "Agent Hospital," a virtual environment where all entities—patients, nurses, and doctors—are represented by autonomous agents with large language model (LLM) capabilities. The introduced MedAgent-Zero methodology allows doctor agents to learn and refine treatment strategies through simulated medical experiences, encompassing both successful and unsuccessful treatments. The system shows remarkable improvement in agents' performance over time. Notably, the artificial intelligence (AI) trained within this simulacrum demonstrates transferable skills, achieving a 93.06% accuracy rate on the MedQA dataset related to major respiratory diseases—suggesting potential real-world medical applications.

DeepSeek-V2 - DeepSeek-V2 is introduced as an advanced Mixture-of-Experts language model with 236B parameters, delivering strong performance with economical training and efficient inference. Outperforming its predecessor (DeepSeek 67B), it saves on training costs and enhances throughput. The model excelled across standard and open-ended benchmarks, after pretraining on an 8.1 trillion token corpus and Supervised Fine-Tuning and Reinforcement Learning optimization. DeepSeek-V2 boasts innovations like Multi-head Latent Attention and MoE architecture for training efficiency. It provides an OpenAI-compatible API and can be utilized with 80GB*8 GPUs in BF16 format for inference.

Granite Code Models: A Family of Open Foundation Models for Code Intelligence - The Granite series, a family of large language models (LLMs) for code, is designed to significantly enhance software development. Supporting 116 programming languages and consisting of models with 3 to 34 billion parameters, it excels in various tasks such as generating code, debugging, and providing documentation. The models have been optimized for enterprise workflows and demonstrate state-of-the-art performance, outperforming existing open-source code LLMs. The series is available under an Apache 2.0 license, enabling both research and commercial use, and aims to facilitate a wide range of development tasks efficiently.

AlphaFold 3 predicts the structure and interactions of all of life’s molecules - DeepMind introduces AlphaFold 3, an advanced AI model designed to predict the structure and interactions of all biological molecules with unprecedented precision, significantly surpassing previous models. While AlphaFold 2 focused on protein structure prediction, the new iteration, AlphaFold 3, extends to analyze a full spectrum of biomolecules, which may revolutionize drug discovery and our understanding of biological systems. The AlphaFold 3 model, featuring a novel Evoformer module and diffusion network, is offered for non-commercial use through the AlphaFold Server, aiming to foster scientific progress in areas such as biorenewable materials, crops resilience, and genomics.

The appearance of 'im-a-good-gpt2-chatbot' and 'im-also-a-good-gpt2-chatbot' - Mysterious AI models, 'im-a-good-gpt2-chatbot' and 'im-also-a-good-gpt2-chatbot,' have appeared following the removal of ‘gpt2-chatbot,’ with speculation linking them to OpenAI. It demonstrated extraordinary capabilities such as coding a full Flappy Bird game in one prompt, solving previously unsolvable physics problems, and showing superior knowledge of certain cultural stories. Some theories suggest it could be an early version of GPT-5 or a fine-tuned GPT-2 model, while a cryptic tweet from Sam Altman hints at OpenAI's involvement. Despite the uncertainty, its enhanced reasoning, math, and coding skills are stirring intrigue in the AI community.