OpenAI claims it's nearing Level 2 on its 5-Level Scale, representing 'human-like reasoning' capabilities

OpenAI Scale Ranks Progress Toward 'Human-Level' Problem Solving

OpenAI has introduced a five-tier system to track its progress toward achieving artificial general intelligence (AGI). Currently at Level 1, where AI interacts conversationally, OpenAI aims to reach Level 2, "Reasoners," with AI capable of basic problem-solving at a doctorate level without tools. Levels progress to AI handling tasks autonomously for days, creating innovations, and functioning at an organizational level. This framework helps clarify OpenAI's safety and development goals as it pursues more advanced AI capabilities.

Sponsor

With OnDemand, you’re always one step ahead of the curve.  Join the waitlist today! https://on-demand.io/contact?ref=mberman

• Sequoia, Nvidia Back Startup Fireworks AI at $552 Million Valuation - Fireworks AI, a two-year-old startup, has raised $52 million in a funding round led by Sequoia Capital, valuing the company at $552 million. The round included investments from Nvidia, AMD, and MongoDB, bringing Fireworks AI's total funding to $77 million. The company, founded by former Meta employee Lin Qiao, offers a platform that allows businesses to access and customize over 100 AI models. Fireworks AI focuses on helping companies adopt generative AI without needing large specialized teams. The startup plans to use the funding to expand its workforce and partnerships with AI companies.

• Watch a robot navigate the Google DeepMind offices using Gemini - Google’s DeepMind Robotics team showcases their robot navigation system using the Gemini 1.5 Pro model. The robot responds to natural language commands and navigates the office environment, demonstrated in a series of videos. It uses "Multimodal Instruction Navigation with demonstration Tours (MINT)" to familiarize itself with the office and hierarchical Vision-Language-Action (VLA) for understanding and reasoning. The robot successfully completed 90% of tasks during interactions with employees.

  

• Microsoft Gives Up Observer Seat on OpenAI Board - Microsoft has relinquished its non-voting observer seat on OpenAI's board, citing significant progress and confidence in the company's direction since Sam Altman's reinstatement as CEO. The decision, effective immediately, was made after discussions with OpenAI, reflecting Microsoft's satisfaction with the newly formed board's performance. This move also alleviates potential antitrust concerns regarding Microsoft's deep involvement with OpenAI. Meanwhile, OpenAI plans to establish a new approach for engaging strategic partners and investors, foregoing board observers altogether.

• OpenAI and Los Alamos National Laboratory Announce Bioscience Research Partnership - OpenAI and Los Alamos National Laboratory (LANL) have partnered to explore the safe application of multimodal AI models in bioscientific research. This collaboration aims to harness AI's potential to accelerate scientific progress while addressing associated risks. The partnership will focus on evaluating AI tools, such as GPT-4o, in performing and troubleshooting laboratory tasks. This initiative underscores the importance of public-private cooperation in advancing innovation and ensuring safety in scientific research, aligning with the U.S. government's directives on trustworthy AI development.

  

• Samsung's Jam-Packed Galaxy Unpacked: Galaxy Ring, Z Fold 6 and All the New Products Announced - Samsung's Galaxy Unpacked event unveiled a suite of new gadgets: the innovative Galaxy Ring, the Z Fold 6 and Z Flip 6 phones, Galaxy Watch 7 series, and Galaxy Buds 3 iterations. The Galaxy Ring, a $400 Android-based smart ring, offers health tracking without a subscription but requires a Galaxy phone for full functionality. The Z Fold 6 features an enhanced hinge, brighter display, and Snapdragon 8 Gen 3 processor, starting at $1,900. Meanwhile, the Z Flip 6, starting at $1,100, boasts a new 50-megapixel camera and longer battery life. The durable Galaxy Watch Ultra, at $650, includes advanced health features and a new processor, while the Galaxy Buds 3 and Pro offer improved audio, ANC, and AI capabilities at $180 and $250, respectively. Samsung promises extensive software support for these devices, with a seven-year update guarantee for its phones.

  

• AMD Looks to 'Move Fast' and Swipe at Nvidia with Its Purchase of Silo AI - AMD has announced its acquisition of Silo AI, Europe's largest private AI lab, for $665 million, marking its most significant AI-focused purchase to date. This strategic move aims to enhance AMD's capabilities in AI software, bridging the gap with Nvidia's established dominance. The acquisition brings 300 AI experts, including 125 Ph.D.s, to AMD, bolstering its ability to deliver end-to-end AI solutions and support the adoption of its MI300 AI accelerators. This partnership is crucial for AMD to leverage the growing AI market and achieve its projected $4 billion revenue.

• Intuit will lay off 1,800 workers and hire new ones to advance its AI ambitions - Intuit is restructuring to prioritize artificial intelligence development, resulting in the termination of 1,800 positions (10% of their workforce) with plans to rehire a similar number with a focus on engineering, product, and customer roles. The company will also close its Boise, Idaho and Edmonton, Alberta offices, impacting 250 employees. The move, according to CEO Sasan Goodarzi, is not for cost-cutting but for redirecting resources to essential areas, notably AI. Intuit has set higher employee performance standards leading to 1,050 departures while reducing its executive team by 10%. Despite reductions, Intuit anticipates headcount growth by the 2025 fiscal year. They are also increasing investments in AI, aiming to advance their offerings like the AI-driven financial assistant, Intuit Assist. Post-announcement, Intuit's stock fell over 3%.

• Humane execs leave company to found AI fact-checking startup - Brooke Hartley Moy and Ken Kocienda, former top employees at Humane, have launched Infactory, an AI-driven fact-checking search engine aimed at enterprise customers like newsrooms and research facilities. Unlike other AI tools, Infactory uses large language models for a natural language interface but ensures factual accuracy by pulling data directly from trusted sources with citations, avoiding the "hallucinations" common in generative AI. The founders emphasize that their departure from Humane, amid its struggles with AI hardware, was not influenced by these challenges but driven by a vision to enhance data reliability in professional environments.

• Chinese self-driving cars have quietly traveled 1.8 million miles on U.S. roads, collecting detailed data with cameras and lasers - In February 2023, Fortune investigated Chinese self-driving cars collecting detailed mapping data on California's roads, triggering concerns about national security and privacy. These autonomous vehicles, belonging to companies partly from China, utilize cameras and sensors to precisely chart their surroundings. While they're legally testing in the U.S., no sufficient government scrutiny exists regarding what data they collect or where it goes. Though no evidence suggests misuse by the Chinese government, the legal framework in China could facilitate data access by state authorities. The lax oversight in the U.S. contrasts starkly with China's stringent regulations, which prevent foreign entities from collecting such data within its borders. The U.S. has begun to raise alarms about the potential espionage risks similar to issues raised with TikTok and Huawei. However, there is a significant lag in regulating data security involving these Chinese vehicles, an aspect critics believe needs urgent attention as autonomous vehicle technology progresses.

  

• Etsy loses its “handmade” and “vintage” labels as it takes on Temu and Amazon - Etsy has updated its policies, introducing four product classifications: "made by," "designed by," "handpicked by," and "sourced by." These categories must be used to describe all items on the platform, providing clearer details to customers on the origin and creation process of the products. Despite these changes, Etsy maintains its ban on reselling items not made by the seller. This move reflects Etsy's mission to preserve the human element in its marketplace, amidst past controversies over the dilution of its handmade ethos and challenges from mass-produced goods and AI-generated content. The new categories aim to ensure transparency and support genuine artisans.

• What is AI? - AI, or artificial intelligence, remains a contentious and multifaceted term, encompassing technologies that enable computers to perform tasks requiring human-like intelligence. Despite its pervasive influence, defining AI elicits diverse and often conflicting perspectives, highlighting issues of trust and understanding. AI's capabilities, ranging from recognizing faces to generating text, prompt debates about its true nature and potential. This discord is intensified by the contrasting views of techno-optimists, who see AI as transformative, and skeptics, who caution against overestimating its abilities. The lack of consensus complicates discussions on AI's role and impact in society, underscoring the need for clearer definitions and informed discourse.

Sponsor

Check out more details and get your own ASUS Vivobook S 15 ➜ https://asus.click/vbs_matthew

• French Startup Bioptimus Releases AI Model for Disease Diagnosis - Paris-based startup Bioptimus has launched H-optimus-0, an AI model trained on hundreds of millions of images to assist in disease research and diagnosis. The model, capable of identifying cancerous cells and genetic abnormalities, is the largest open-source pathology model available. Bioptimus aims to enhance transparency and accelerate medical advancements. Despite concerns about AI in healthcare, the company emphasizes that H-optimus-0 is just the beginning of its broader vision. Founded in February with $35 million in seed funding from investors like Bpifrance and Xavier Niel, Bioptimus seeks to push the boundaries of AI in medicine.

• Hewlett Packard Enterprise to Deliver AIST’s Next-Generation Supercomputer Powered by NVIDIA for Generative AI - Hewlett Packard Enterprise (HPE) announced it will build the next-generation supercomputer, ABCI 3.0, for Japan's AIST. Powered by NVIDIA H200 Tensor Core GPUs, this supercomputer aims to support large generative AI models, advancing research and development across various sectors. The collaboration highlights HPE and NVIDIA's commitment to enhancing AI capabilities and providing cloud services for public and private entities.

• xAI Appears to Confirm Ended Talks With Oracle Over Expanded AI Chips Agreement - Elon Musk's xAI has confirmed it ended discussions with Oracle about expanding their agreement to rent Nvidia chips. This decision coincides with xAI's efforts to build a supercomputer with 100,000 Nvidia GPUs, positioning it as the world's most powerful training cluster. xAI aims to outpace competitors in AI development, relying on its new data center in Memphis, Tennessee. This move follows Musk's continued push to enhance xAI's capabilities, including hiring top engineers and launching the Grok chatbot.

  

Awesome Research Papers

• A Survey on Mixture of Experts - This paper describes a comprehensive survey on large language models (LLMs), particularly focusing on the mixture of experts (MoE) technique. LLMs' capabilities are attributed to their size, diverse datasets, and computational resources used in training. MoE stands out for enabling model scaling with minimal additional computation. The survey fills the gap in systematic MoE literature review by introducing MoE's structure, proposing a taxonomy, discussing core designs, and listing open-source resources. It also examines MoE's applications and future research directions, with updates shared via an established resource repository.

• Assessing ASR performance with meaning preservation - This research focuses on assessing the comprehensibility of automated speech recognition (ASR) systems beyond traditional metrics like word error rate (WER). It posits that meaning preservation is essential, especially for atypical speech, which suffers from high WER. The authors use a large language model (LLM) to predict whether ASR transcripts maintain the intended meaning. Through Project Euphonia, they collected a corpus of disordered speech and trained a classifier on top of Google's Gemini (a smaller, efficient LLM), achieving high accuracy in meaning preservation assessment. The system showed strong performance in English and generalized well without additional training to Spanish and French, important for assistive technologies like Project Relate.

• UltraEdit: Instruction-based Fine-Grained Image Editing at Scale - UltraEdit is a newly presented, large-scale dataset with around 4 million samples for instruction-based image editing. It aims to surpass previous datasets by combining creative inputs from large language models and human raters, employing real images for diverse and less biased content, and utilizing enhanced region-based annotations. The dataset has been shown to improve performance in diffusion-based editing baselines, achieving top results on established benchmarks.

• Internet of Agents: Weaving a Web of Heterogeneous Agents for Collaborative Intelligence - The paper introduces the Internet of Agents (IoA), a new multi-agent framework designed to overcome the limitations of existing systems in integrating capable third-party agents and simulating distributed environments. IoA features an agent integration protocol, a communication architecture akin to instant messaging, and adaptable mechanisms for team formation and dialogue management. It has proven its effectiveness over current state-of-the-art solutions through rigorous testing across various tasks, demonstrating enhanced collaboration between diverse large language model (LLM)-based agents.

AI-MO/NuminaMath-7B-TIR - NuminaMath 7B TIR is a 7 billion parameter language model designed specifically for solving math problems using a technique known as tool-integrated reasoning (TIR). The model excels in competition-level mathematics, having achieved a 29/50 score in the AI Math Olympiad. It is the result of a two-stage supervised fine-tuning process: initially using a dataset of natural language math problems with templated solutions, and subsequently with synthetic tool-integrated reasoning data based on Microsoft's ToRA format. However, its efficacy is limited to mathematics and it may struggle with more complex problems and lacks visual processing capabilities necessary for geometry problems.

Introducing EVE: The New Encoder-Free Vision Language Model - The new encoder-free Vision Language Model (VLM) called EVE has been launched, offering support for arbitrary image resolutions similar to Fuyu-8b, but surpassing it in benchmark performance. Trained on 35 million publicly accessible data, EVE is designed for general-purpose VLM tasks, unlike Fuyu's UI-specific applications. EVE's transparency in training data and methods, alongside its MIT License, promotes reproducibility and further research advancements in the AI community.

Ocean - Ocean is an open-source framework licensed under MIT, designed for high-performance development of Computer Vision and Mixed Reality applications across multiple operating systems. With a core written in C/C++, it enables the creation of efficient and smooth-running native applications compatible with a range of devices. Ocean offers a variety of demo applications and invites developers to use these for inspiration or as a learning resource. Notably, developers can access camera feeds on Quest devices using external cameras.

Odyssey - Odyssey introduces a Hollywood-grade visual AI aimed at revolutionizing storytelling through technology. They aim to counteract current trends of low-quality AI content by handing over sophisticated AI tools to professional storytellers, allowing them to generate and direct high-quality, visually stunning videos, and maintain storytelling prowess. This breakthrough technology offers four generative models to control geometry, photorealistic materials, lighting, and motion—offering unprecedented detail and creativity for movie, TV, and video game production. These tools promise to integrate seamlessly with existing Hollywood and gaming production workflows, offering fine-tuned control and the potential to imagine, generate, and iterate worlds with pro-grade version control.

Ex-Googler joins filmmaker to launch DreamFlare, a studio for AI-generated video - DreamFlare AI, co-founded by former Google employee Josh Liss and filmmaker Rob Bralver, has launched a platform to help creators produce and monetize AI-generated short-form content. Utilizing third-party AI tools, DreamFlare acts as a studio where creators collaborate with professional storytellers. The platform offers two main content types: "Flips," comic book-style stories with AI-generated clips, and "Spins," interactive choose-your-own-adventure films. Despite concerns about AI's impact on jobs, DreamFlare aims to democratize storytelling and provide new revenue opportunities for creators through subscriptions, ad revenue, tips, and merchandise sales.

Announcing Tessl, the AI Native Development Startup — Tessl - Tessl is spearheading a new approach to software development for the AI era, termed "AI Native Software Development." This method integrates AI deeply within the development process, rather than using it merely as a supplemental aid to existing methodologies. While the industry has embraced AI-assisted tools for boosting productivity and automating tasks like code completion, testing, and documentation, Tessl envisions a more profound transformation. They are exploring how AI can fundamentally change the roles of software developers and the creation process itself, raising questions about the future of software development in a world where AI is not an add-on, but a core component of every aspect of development.

AWS + Scale Partner to Accelerate Generative AI Adoption - The announcement highlights a significant strategic partnership between Scale and Amazon Web Services (AWS) aimed at boosting generative AI (GenAI) adoption in enterprise and public sectors. This collaboration addresses common deployment hurdles by establishing a platform for AWS customers to customize and trust GenAI models using proprietary data. The partnership offers the Scale GenAI Platform via AWS Generative AI Innovation Center or AWS Marketplace, enhancing GenAI app evaluation and model fine-tuning through Amazon Bedrock. Additionally, for government entities, the Donovan AI platform, tailored for federal security compliance, is accessible on AWS Marketplace, promising to enhance productivity and mission outcomes.

Poe Introduces Previews for Interactive Web Applications - Poe has launched a new feature called Previews, enabling users to interact with web applications directly within chats. This feature supports LLMs such as Claude 3.5 Sonnet, GPT-4o, and Gemini 1.5 Pro, and allows users to share interactive outputs via dedicated links. Previews facilitate the creation of custom web applications, including games, data visualizations, and interactive animations, without requiring programming skills.