ElevenLab's new AI breakthrough

Good morning. It’s Wednesday, October 11th.

Did you know: 17 years ago, Google announced their purchase of YouTube for $1.65 billion in stock?

In today’s email:

  • AI Technology Innovations

  • Corporate AI Developments & Acquisitions

  • AI in Agriculture & Environmental Management

  • Analyses and Predictions about AI Technology

  • 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 trending AI news stories

AI Technology Innovations

ElevenLabs Launches Voice Translation Tool to Break Down Language Barriers for Content This tool instantly translates spoken content into multiple languages while preserving the original speaker’s voice. It enables creators, educators, and media companies to reach global audiences with ease, significantly expanding content accessibility across various mediums. This breakthrough advances content accessibility and supports over 20 languages.

Try it here on ElevenLabs’ website.

Here’s a demo video of “AI Dubbing” in action:

StreamingLLM Lets AI Models Running Indefinitely with One Token AI researchers have introduced StreamingLLM, a framework designed to maintain the quality of responses generated by Large Language Models during extended conversations, even when user prompts exceed the trained token limit. This development addresses the performance dip that occurs when LLMs are used for prolonged conversations. StreamingLLM is a collaboration between Meta, MIT, and Carnegie Mellon University, which utilizes “attention sinks” to improve LLM performance. It enables LLMs to handle text of infinite length without fine-tuning and reduces attention deflection when users exceed the token limit, paving the way for human-like chatbot responses during long discussions.

Mendel Launches Hypercube, an AI-Copilot for Real-World Data Applications Hypercube facilitates analysis of both structured and unstructured patient data, enabling quick insights and complex cohort analysis through plain English queries. Unlike traditional AI methods, Hypercube uses a hybrid approach blending large language modeling with symbolic AI, making it more suitable for clinical applications. The technology’s potential for clinical trial screening and data analysis efficiency has garnered attention from researchers at the University of Pennsylvania School of Medicine. You can explore this project on Mendel’s website here.

Adobe Firefly can now generate more realistic images with Firefly Image 2 Model. This new model significantly improves the rendering of human images, focusing on facial features, skin, body, and hands. It was trained on more recent images from Adobe Stock and other sources, increasing its dataset for better user understanding. This model will be available on the Firefly web app, and in the near future, in Creative Cloud apps like Photoshop. Adobe emphasizes that its approach to generative AI is focused on generative editing.

MagicSchool brings 50+ AI tools to modern educators, leveraging generative AI to support teachers in planning lessons. It has already gained over 150,000 users since its launch. The AI tool offers more than 50 different tools for educators, allowing them to adapt texts for various reading levels, create worksheets, generate science experiments, explain concepts, and more. MagicSchool aims to ease the burden on teachers, especially in light of the 30,000 teacher shortage in the U.S. The tool is currently free, with a premium version expected to launch soon.

Corporate AI Developments & Acquisitions

AMD acquires open-source AI software pioneer Nod.ai to fortify AI capabilities As AI chip technology rapidly expands, AMD aims to bolster its presence in this burgeoning market. The acquisition aligns with AMD’s strategy of fostering an open software ecosystem to simplify AI adoption and expand into the growing AI industry. Nod.ai’s software, like the SHARK Machine Learning Distribution, will enhance AMD’s capacity to offer high-performance AI models tailored for AMD hardware. This strategic move is part of AMD’s ongoing investment in AI technology to compete with companies like Nvidia and Intel.

Microsoft Teams Gets Free Mesh Toolkit to Help Develop Immersive Spaces and is set to enter its public preview phase for Microsoft Mesh, a service that enables immersive virtual interactions within Teams meetings. This toolkit empowers developers to create and customize virtual environments with various tools, including graphics, physics, interactivity, visual scripting, and more. To utilize Microsoft Mesh, developers would require a Teams Premium License. These immersive spaces can be designed in Unity and uploaded to the Mesh Portal for use by other Mesh users.

AI in Agriculture & Environmental Management

AI drones to help farmers optimize vegetable yields Researchers at the University of Tokyo have demonstrated an AI-powered system for optimizing vegetable harvest times using drones. These low-cost drones, equipped with specialized software, capture and analyze data about young plants to predict their growth characteristics. Accurate harvest predictions can reduce income loss for farmers by up to 20.4%. The system relies on deep learning to process the image data collected by drones and provide easy-to-understand visual data for farmers, making it a valuable tool for optimizing crop yields.

AI-powered earthquake forecasting proves to be a success Researchers at the University of Texas at Austin have achieved a 70% accuracy rate in predicting earthquakes using an AI algorithm. During a seven-month trial in China, the AI was trained to detect statistical anomalies in real-time seismic data and historical earthquake records. The AI achieved this milestone by predicting 70% of earthquakes a week in advance, while 14 forecasts came true within 200 miles of their estimated locations.

Analyses and Predictions about AI Technology

'Overhyped' generative AI will get a 'cold shower' in 2024, analysts predict Generative AI, exemplified by models like OpenAI’s ChatGPT and Google Bard, relies on a huge amount of computing power and chips to run complex mathematical models. With increasing costs, smaller developers may find it prohibitively expensive. Additionally, CCS Insight predicts that AI regulation in the European Union will face obstacles due to rapid AI advances, leaving room for self-regulation. The firm also forecasts labels on AI-generated content and arrests for AI-based identity fraud in 2024. Personally, I think someone has to play Devil’s advocate with the AI hype, but I disagree with the “cold shower” theory.

5 new AI-powered tools from around the web

CallZen.AI is a conversational AI tool designed for customer success. This SaaS solution offers conversational insights, AI-powered analytics, multilingual transcription, sentiment analysis, automated compliance management, and more for contact centers. It empowered businesses to make each dialogue impactful with valuable insights from conversation data.

SEO Quickr is an AI-powered SEO tool that enhances website visibility and search engine rankings. With in-depth SEO analysis, keyword discovery, and AI-generated SEO content, it simplifies the process of improving online presence.

Ceacle Pipeline offers automation for creators, enabling the creation of custom pipelines manually or with AI assistance. It streamlines various tasks, from generating inspiration boards to batch editing and creating multiple AI images in a single step. A valuable tool for boosting productivity and creativity in design and more.

AppManager simplifies IT for startups by harnessing the power of AI to streamline user provisioning and app management. This affordable and effortless solution empowers startups to focus on growth while leaving IT management in capable hands.

Pelery simplifies the process of repurposing Substack newsletter content for various social media platforms, boosting engagement and growth across your entire newsletter archive. This AI-powered tool offers a solution for authors who want to promote their content without the hassle, and it’s currently free to use.

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

The research conducted by Google and California State University researchers examines the anticipated impact of generative AI on various knowledge industries. Involving 54 participants from seven industries, it reveals that knowledge workers largely envision generative AI as a tool for performing menial tasks under human supervision, rather than causing significant disruptions. These participants anticipate generative AI amplifying forces like deskilling, dehumanization, disconnection, and disinformation within their industries. The study aims to contribute valuable insights for the HCI (Human-Computer Interaction) community while providing a nuanced understanding of how generative AI is expected to influence knowledge work.

The Autonomous Cognitive Entity (ACE) model presents a comprehensive framework for building ethical and intelligent autonomous agents. Comprising six layers, this model integrates moral reasoning, value alignment, cognitive control, and task execution. It draws inspiration from the OSI model and leverages insights from diverse fields, including neuroscience, philosophy, and AI ethics. The ACE framework aims to bridge the gap between AI capabilities and ethical reasoning, enabling AI systems to make ethically sound decisions while operating autonomously. By unifying these critical elements in a layered architecture, it sets a new standard for the development of capable and ethical AI agents.

In a recent study, a novel cognitive science-inspired protocol called CogEval was introduced for a comprehensive evaluation of cognitive abilities in large language models (LLMs). This protocol was applied to assess cognitive maps and planning abilities with eight LLMs, which included widely used models. The study scrutinized LLM performance on various tasks, examining their adaptability, planning, and understanding of task structures. Results revealed that while LLMs exhibited proficiency in simpler planning tasks, systematic evaluation uncovered significant limitations, including generating invalid trajectories and getting trapped in loops. These findings cast doubt on the idea of LLMs having inherent out-of-the-box planning capabilities.

The paper examines the pivotal role of length in Reinforcement Learning from Human Feedback (RLHF) settings, with a particular emphasis on enhancing model “helpfulness.” It reveals that optimizing for length significantly influences RLHF’s reported improvements. Length increases are a prevalent occurrence during Proximal Policy Optimization (PPO), and they correlate with improved reward model scores. Despite interventions, such as PPO adjustments and alterations to preference data, the inclination to optimize for length persists. Length biases originating from imbalanced preference data contribute to this phenomenon, underscoring the challenge of disentangling length from reward optimization in RLHF.

The paper presents DeepSpeed-Ulysses, an innovative methodology designed to facilitate the efficient and scalable training of large language models, particularly those requiring the processing of exceptionally long sequences. It addresses the growing need for longer context in conversational AI, scientific research, and healthcare. DeepSpeed-Ulysses focuses on sequence parallelism and minimizes communication inefficiencies, supporting different types of attention mechanisms and integrating with ZeRO-3 for memory optimization. Through rigorous evaluation, it showcases impressive scalability and increased throughput compared to existing solutions, making it a promising advancement for training extensive models with extended sequences.

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.

Peace out.