StabilityAI co-founder's $500M mistake

Author

AI Breakfast
July 14, 2023

Good day. It’s Friday, July 14th.

Did you know: Cyrus Hodes, co-founder of Stability AI, is suing the company alleging that he was tricked into selling his 15% stake for only $100?

In today’s email:

  • Stability AI Releases Stable Doodle, a Sketch-to-Image Tool

  • FTC Investigates OpenAI Over ChatGPT's 'Reputation Harm'

  • Stability AI Co-Founder Sues Over Undervalued Stake Sale

  • Elon Musk's xAI Company Formed for Universal Understanding

  • NVIDIA Invests $50M in Recursion for AI Drug Discovery

  • Meta to Launch Commercial AI Model to Rival OpenAI and Google

  • Google's Art Project: Animated Bird and AI Playing Cello

  • Intel Offers Customized AI Processors to China

  • Google DeepMind, OpenAI, and Academics Propose Global AI

  • Governance

  • Objaverse-XL Dataset Enhances AI Training

  • 5 New AI Tools

  • Latest AI Research Papers

You read. We listen. Let us know what you think of this edition by replying to this email, or DM us on Twitter.

Today’s edition is brought to you by:

CoPilot AI

CoPilot has launched the first ChatGPT plugin for car shopping, providing actionable AI-powered analysis & rankings of every car with data from across the internet and CoPilot proprietary databases.

Your CoPilot Intelligent Agent brings actionable AI to your car shopping experience: Searches every dealer to identify every matching car for sale in your area based on your specific needs  

AI-powered analysis & rankings harnessing data from across the internet in combination with proprietary CoPilot databases Unlike other major car shopping sites & apps which return results based on which dealers pay them  

Two New Tools: ChatGPT Plugin & Model Discovery Chat Tool Take the grind out of car shopping, let AI find your next car for you!

Today’s trending AI news stories

Stability AI releases Stable Doodle, a sketch-to-image tool: Stability AI has introduced Stable Doodle, a sketch-to-image service that uses its Stable Diffusion model to generate visually appealing images based on user sketches. The tool, powered by OpenAI’s Stable Diffusion XL and TenCent’s T2I-Adapter, offers more precise control over image generation compared to sketch-to-image AI tools. Stable Doodle aims to be accessible to both professional and novices, allowing anyone with basic drawing skills to create high-quality original images. The service is currently available through ClipDrop, which Stability acquired earlier this year.

FTC reportedly looking into OpenAI over ‘reputation harm’ caused by ChatGPT: The Federal Trade Commission (FTC) is reportedly investigating OpenAI over potential ‘reputation harm’ caused by its ChatGPT conversational AI. The FTC is exploring whether the AI system made false or misleading statements about individuals. The investigation signals that the FTC is taking action beyond issuing warnings to the AI industry.

Stability AI Co-Founder Hodes Says He Was Duped Into Selling Stake for $100: Cyrus Hodes, co-founder of Stability AI, is suing the company and its CEO Emad Mostaque, alleging that he was tricked into selling his 15% stake in the company for only $100. Hodes claim that Mostaque convinced him that the company was essentially worthless, but three months later, Stability AI secured $101 million funding round with a post-money valuation of $1 billion. Hode’s shares, if retained, could now be worth $500 million.

Elon Musk’s new xAI company launches to ‘understand the true nature of the universe: xAI is a newly formed company led by Elon Musk with the goal of understanding the true nature of the universe. The team comprises experts from renowned organizations such as DeepMind, OpenAI, Google Research, Microsoft Research, Tesla, and the University of Toronto. They already have made invaluable contributions in the field, including the development of widely used methods and breakthroughs like AlphaStar and GPT models. While separate from X Corp, xAI will collaborate closely with X (Twitter), Tesla, and other companies. The team is actively recruiting engineers and researchers to join their Bay Area-based staff.

Nvidia deepens bets on AI in drug discovery with Recursion investment: NVIDIA plans to invest $50 million in Recursion to enhance the training of its AI models for drug discovery. Recursion will use its extensive biological and chemical datasets to train AI models on NVIDIA’s cloud platform. The collaboration aims to accelerate the development of AI-driven drug discovery solutions and could lead to the licensing of these models to biotech firms through BioNeMo. While NVIDIA’s investment propelled Recursion’s shares to surge, the potential of AI models in revolutionizing drug discovery continues to be explored.

Meta to release commercial AI model in effort to catch rivals: Meta is preparing to launch a commercial version of its AI model to rival Microsoft-backed OpenAI and Google in the race for generative AI. The forthcoming model will be more widely accessible and customizable for businesses, enabling them to develop tailored software using Meta’s technology. By adopting an open-source approach, Meta aims to catch up with competitors and enhance its AI capabilities by leveraging user input. Meta’s vice-president of AI research, Joelle Pineau, emphasized the company’s intention to incorporate these models into its own products while retaining intellectual property rights.

Google’s latest art project recruits an animated bird — and AI — to play the cello: Google’s latest art project featured an animated bird playing with cello with the help of AI. The project showcases Google’s focus on generative AI, despite competition from other tech giants.

Intel offers customized AI processors to China, backed by major Chinese AI server providers: Intel has announced the launch of its deep learning processor, Habana Gaudi 2, in the Chinese market. The processor is designed to accelerate AI training and inference tasks and is expected to be featured in server products by Chinese AI server giants such as Inspur, New H3C, and xFusion. Intel aims to compete with Nvidia in the Chinese market by providing an alternative option. The company plans to further update its data center product roadmap and integrate high-performance AI chips with GPUs for a comprehensive next-generation GPU product by 2025.

Google DeepMind, OpenAI, and Leading Academics Propose International Institutions for Global AI Governance: Google DeepMind, OpenAI, and leading academic institutions have collaborated to propose international institutions for global AI governance in a new white paper. The paper underscored the need for shared international standards to ensure responsible development and use of advanced AI systems. It suggests four potential institutional models, including a Commission on Frontier AI, an Advanced AI Governance Organization, a Frontier AI Collaborative, and an AI Safety Project. The paper also discussed challenges and open questions regarding funding, jurisdiction, cooperation, representation, and conflicts of interest.

Objaverse-XL brings 10 million 3D objects to generative AI training: Objaverse, a dataset of over 10 million 3D objects, has been used to train the Zero123-XL model, showcasing its strong zero-shot generalization across different modalities. This dataset is anticipated to improve the capabilities of AI models in 3D applications, including augmented and virtual reality.

🎧 Did you know AI Breakfast has a podcast read by a human? Join AI Breakfast team member Luke (an actual AI researcher!) as he breaks down the week’s AI news, tools, and research: Listen here

5 new AI-powered tools from around the web

Beebee AI is a platform for in-depth financial analysis of public companies. Analyzes earnings call transcripts and financial data, providing concise insights. Its features include automatic analysis, key numbers, market sentiment, as well as generating suggested questions based on meeting records. Free access for S&P 500 and more.

Tavily is an open-source autonomous agent for detailed and unbiased online research. It provides comprehensive research reports on any topic, optimizing performance with parallelized work for increased speed and stability. It’s free to use.

Nekton is an AI-powered automation platform that enables users to automate tasks using plain English descriptions. It integrates with thousands of services, generates automation code, and runs it in the cloud. The platform offers easy customization, forms, scheduling and no maintenance requirements. Ideal for automating small, specific tasks efficiently.

Bizway is an AI-powered business planning platform that guides users in transforming ideas into comprehensive business plans. It offers step-by-step assistance, including competitor analysis, name brainstorming, and financial forecasting.

Ortus is an AI-powered Chrome extension designed to enhance learning by providing interactive video viewing experiences. Users can ask questions, receive relevant answers, access high-quality summaries, seamlessly transfer key insights to Notion, and engage with a supportive Discord community for feedback and support.

arXiv is a free online library where scientists share their research papers before they are published. Here are the top AI papers for today.

SayPlan is a new approach to large-scale task planning for robotics using 3D scene graphs and language models. It addresses the challenge of grounding language models in expansive environments by exploiting the hierarchical nature of 3D scene graphs and integrating classical path planning and iterative replanning techniques. The scalability and effectiveness of SayPlan are demonstrated through evaluations in large-scale environments, showcasing its capability to generate feasible long-horizon task plans from abstract and natural language instructions. The proposed approach provides a promising solution for robotics task planning in complex and multi-room environments, contributing to the field of robotics and language understanding.

This paper from Stanford researchers explores the use of human video demonstrations for training robotic manipulation policies. By leveraging eye-in-hand cameras and a simple image masking technique, they bridge the domain gap between human and robot data without explicit domain adaptation methods. The study demonstrates significant improvements in generalization across real-world tasks, enabling robots to generalize to new environments and tasks not seen in the robot demonstration data. The approach offers a cost-effective and scalable solution for training vision-based robotic manipulators.

POLYLM is an open-source multilingual LLM designed to address the limitations of LLMs primarily focused on high-resource languages. Developed by Alibaba Group’s DAMO Academy, POLYLM is trained on dataset of 640 billion tokens, with a focus on non-English languages. It employs a curriculum learning strategy to improve the transfer of knowledge from English to other languages and introduces a multilingual self-instruct method for model fine-tuning. POLYLM outperforms other open-source models on multilingual tasks while maintaining comparable performance in English. The model, instruction data, and multilingual benchmark are available for further research.

NaViT (Native Resolution ViT) is a vision transformer model that introduces Patch n’ Pack, a technique that allows processing inputs of arbitrary resolutions and aspect rations. Unlike traditional computer vision models that resize images to fixed resolutions, NaViT utilizes sequence packing during training, treating images as sequences of patches. This approach offers flexibility in model usage, improved training efficiency, and enhanced performance on various tasks such as image and video classification, object detection, and semantic segmentation. NaViT marks a departure from the standard input and modeling pipeline used in computer vision models and shows promise for the future of vision transformers.

DNAGPT is a pretrained tool designed for DNA sequence analysis tasks. It utilizes pre-training on reference genomes from various species and incorporates a symbolic language to handle different downstream tasks. The model can process DNA sequences and numbers simultaneously, improving classification, regression, and generation tasks. By unifying diverse data types and task paradigms, DNAGPT expands the application of pretrained models in biology. It provides a comprehensive solution for leveraging large-scale models in DNA sequence analysis, opening doors to new discoveries in the field.

Thank you for reading today’s edition.

Your feedback is valuable.


Respond to this email and tell us how you think we could add more value to this newsletter.

Attending Ai4 this year? We will be!