Welcome to this week’s exploration of AI's transformative power across multiple dimensions—from enhancing software development and web navigation to revolutionizing genetic medicine and autonomous driving. We delve into cutting-edge AI applications that predict, adapt, and evolve, not only reshaping industries but also challenging our conceptual boundaries. Below, some of the most compelling developments that demonstrate AI’s expansive impact on our world from this week!
Top AI news this week:
🐎 Introducing WebLlama - this new AI Agents will autonomously browse the web by following your instructions - https://github.com/McGill-NLP/webllama
🎬 The integration of Adobe Firefly into Premiere Pro shows us a future where the differentiation between real and AI-generated videos... becomes really difficult:
💻 AI will not only solve, but even predict and repair bugs for developers directly in the IDE: https://research.google/blog/safely-repairing-broken-builds-with-ml/
Case study of the Week
💊 AI in Pharmaceutical Industry - case study on cooperation of Moderna & OpenAI. Moderna leverages OpenAI’s technology to transform various aspects of their operations, from clinical trials to regulatory compliance, testing AI’s potential in the pharmaceutical industry.
We’re looking at every business process— from legal, to research, to manufacturing, to commercial—and thinking about how to redesign them with AI. - Stéphane Bancel, CEO of Moderna
Within two months of the ChatGPT Enterprise adoption:
Moderna had 750 GPTs across the company.
40% of weekly active users created GPTs.
Each user has 120 ChatGPT Enterprise conversations per week on average.
Interesting Food for Thought
📞 CEO of Tata Consulting: The End of Call Centers already within a Year?
The CEO of Tata Consulting suggests that AI advancements could dramatically reduce the need for traditional call centers within a year (!), signaling a major shift in how businesses manage customer interactions. https://www.ft.com/content/149681f0-ea71-42b0-b85b-86073354fb73
🎨 Impact of AI on creativity: AlphaGo is turning 10 years old 🎂!
What is the impact of AlphaGo on Go after 10 years? → Professional Go players not only adapted but also enhanced their creativity after competing with AI.
🚗 Autonomous Vehicles: San Francisco vs. India - You grow with your challenges!
San Francisco faces a growing challenge in the development of self-driving cars due to very strict regulatory restrictions (https://www.slowboring.com/p/self-driving-cares-are-underhyped). Meanwhile, India's unique driving conditions provide a rigorous testing ground for AI-driven autonomous vehicle technology - meet the company Swaayatt Robots, mastering the challenges! https://spectrum.ieee.org/india-self-driving-car
🖼️ A neural network.. that creates neural networks! Maybe Diffusion Models can do more, as we thought? A groundbreaking approach using diffusion models to autonomously generate high-performing neural network parameters, showcasing AI’s potential to self-improve and extend applications beyond traditional boundaries.
News from the Industries
🏦🤖 New Reasearch on AI for RPA by …. JP Morgan!
The automation of business workflows is often a complex, time-consuming, slow and expensive process. JP Morgan is investing in AI driven development of these Robotic Process Automation (RPA) systems on demand. JP Morgan introduces FlowMind, an innovative approach to RPA using large language models to automate workflows, ensuring data integrity and streamlining user interactions. https://arxiv.org/abs/2404.13050
🧬 AI and Human Genome Editing: Research demonstrates AI's capability to design gene editors, successfully editing DNA in human cells. A significant leap in genetics and medicine. Will we see a “prompt-to-new-medications” someday? https://www.profluent.bio/blog/editing-the-human-genome-with-ai
👓 Vision Model integrated in your glasses: Ray-Ban Meta smart glasses, now powered by Meta AI, offering real-time, hands-free information about your surroundings. https://about.fb.com/news/2024/04/new-ray-ban-meta-smart-glasses-styles-and-meta-ai-updates/
AI x Robotics
🤖 MenteeBot can complete tasks end-to-end from verbal commands, including navigation, locomotion, scene understanding, object detection, and more.
Not nearly as sleek as Figure, but still cool to see more humanoid robot startups. Check out this video of Figure as a reminder for comparison:
New Tools
📹 Simulon is in public beta! This is a tool which rapidly integrates 3D models into real footage, reshaping video editing. Guess, what on this vide is NOT real: https://x.com/javilopen/status/1781992198587760866
New Models
📱 Microsoft unveils Phi-3, a small but potent language model that competes with giants like GPT-3.5, designed to operate efficiently on mobile devices. https://arxiv.org/abs/2404.14219
📱 Apple presents OpenELM - inference and fine-tuning on iPhone?? More accurate LLMs with fewer parameters and fewer pre-training tokens. It is an open language model, utilizing a layer-wise scaling strategy for efficient parameter allocation and enhanced accuracy, while also enabling inference and fine-tuning on Apple devices through integration with the MLX library. https://arxiv.org/abs/2404.14619
❄️ Snowflake Arctic: Setting a New Standard in LLM Efficiency
Snowflake introduces Arctic, an open-source language model that combines efficiency with powerful performance, using far less compute to set new benchmarks in AI development. https://www.snowflake.com/blog/arctic-open-efficient-foundation-language-models-snowflake/
⚗️ New Frameworks and Research work
📝 An interesting paper on the theoretical part: 🌀 Self-Evolution in Large Language Models, about the potential for large language models to autonomously evolve, drawing parallels with human learning processes and highlighting a framework that could enable continual learning and improvement in AI systems. https://arxiv.org/abs/2404.14387
🕸️ AutoCrawler: Evolving Web Crawling: AutoCrawler melds LLMs with web crawlers, optimizing web task automation across varied environments. https://arxiv.org/abs/2404.12753
🎨 ByteDance teach LLMs Grafic Design and Layers!
ByteDance unveils Graphist, utilizing large multimodal models to teach LLMs advanced graphic design skills, explaining how to work with layers in graphics. https://github.com/graphic-design-ai/graphist
🏫🤔 Theorem co-pilot: How to teach LLMs to prove real theorems? Lean Copilot introduces large language models to Lean, significantly advancing proof automation by suggesting tactics and searching for proofs to streamline the theorem proving process. https://github.com/lean-dojo/LeanCopilot
🔗 Graph Machine Learning and LLMs: Exploring the intersection of Graph Machine Learning and LLMs, this survey highlights how integrating graph data can enhance language models, offering new possibilities in various analytical applications. https://arxiv.org/abs/2404.14928
📚 Exploring AI's Capabilities in Complex Tasks: AI agent architectures, focusing on their enhanced abilities in reasoning, planning, and tool use, suggesting a profound shift in how AI systems could tackle complex, multi-faceted tasks in the future. https://arxiv.org/abs/2404.11584
🧠 Perfecting Problem Solving in AI - DUP prompting technique pushes the limits of problem-solving in AI, achieving over 97% accuracy on the GSM8K benchmark by mimicking human problem-solving methods, representing a significant improvement over traditional approach. https://arxiv.org/abs/2404.14963
👨💻 NExT: Enhancing LLMs' Understanding of Code
A new teaching method for LLMs, NExT enhances their ability to reason about code execution, aiming to bridge the gap between AI understanding and practical programming tasks. https://arxiv.org/abs/2404.14662
📊 DataTune: Refining Synthetic Data for NLP. Task-specific data is not available for many use cases, and manually curating task-specific data is labor-intensive. DataTune is a method that transforms publicly available datasets into highly specific, task-oriented data, addressing the scarcity of tailored datasets for NLP applications. https://arxiv.org/abs/2404.14361
Thank you for joining us on this week’s journey through the exciting developments in AI. If you found these insights valuable, feel free to share your thoughts and stay connected for more updates next week!