Weekly AI - 25 Dec 2023
Hello readers! Welcome to Weekly AI. In this issue, we'll learn about major AI art advancements with Midjourney's new update, Tesla's humanoid robot improvements, OpenAI's content partnership, and more.
I always appreciate your thoughts and questions. Please keep sending feedback so we can have an open dialogue about artificial intelligence and its impacts. For now, let's jump in!
This newsletter is also available in Spanish and Catalan.
Midjourney Unveils Groundbreaking V6 Update
Midjourney has unveiled its highly anticipated V6 update on December 22nd, 2023, elevating AI-generated art to unprecedented levels. The new version introduces the ability to produce strikingly realistic and detailed images with life-like accuracy. Users can now also render coherent text as part of their designs, resolving a major limitation in previous iterations. Additionally, V6 brings enhanced natural language understanding and prompt engineering for more control over outputs. Early user creations showcase the new dimensions unlocked across landscapes, textures, typography and more.
Tesla Unveils Next-Generation Humanoid Robot Prototype
Tesla released a video showcasing the latest iteration of its humanoid robot prototype, Optimus Gen 2. Compared to the shaky robots displayed just one year ago, Tesla claims this version can walk 30% faster, is 10kg lighter, and demonstrates improved balance, dexterity, and object manipulation. Though unverified, Tesla says “everything in this video is real, no CGI” - meaning the robot’s movements are not simulated or enhanced digitally. While still a prototype under development and not ready for real-world application, the sneak peek highlights Tesla’s rapid progress towards CEO Elon Musk’s vision of mass-producing affordable humanoid robots for mundane tasks people “don’t want to do.”
OpenAI Partners with Axel Springer to Enrich AI Tools with News Content
OpenAI has announced a partnership with global publishing giant Axel Springer to integrate recent news content into its ChatGPT tool. Launching this month, ChatGPT will provide users with summaries and links to articles from Axel Springer's media properties such as Politico and Business Insider. The collaboration aims to give ChatGPT users access to authoritative, up-to-date information while supporting journalism's business model. Both companies stated this partnership represents a meaningful step in leveraging AI to enhance content experiences and sustain quality journalism.
EU Reaches Milestone Agreement on AI Act, Paving Way for Landmark Legislation
Lawmakers in the EU reached a provisional agreement on the proposed Artificial Intelligence Act (AI Act) after intense negotiations. This comprehensive set of rules, anticipated to be the world's first, will govern AI use in Europe and could serve as a model for other countries looking to regulate AI. The agreement establishes obligations for "high-risk" AI systems, like risk assessments and transparency requirements, and gives citizens the right to file complaints and receive explanations about certain AI-powered decisions affecting them. Fines for violations range based on the offense and company size. While some details are still being finalized, the landmark law likely won't take effect until 2025 at the earliest.
OpenAI Expands Safety Efforts, Gives Board Power to Veto Risky AI
OpenAI announced changes to its internal safety processes intended to mitigate potential harms from advanced AI systems under development. A new cross-functional "Safety Advisory Group" will evaluate models and make recommendations, while the Board of Directors has been granted veto power over releasing risky AI. Although details of the updated "Preparedness Framework" remain vague, OpenAI states that models deemed "high risk" cannot be deployed and those with "critical" risks will not be developed further. This move comes on the heels of recent leadership changes and seeks to reassure observers that safety is a priority. However, transparency and independent oversight mechanisms are still lacking.
NotebookLM - Your Personalized AI Research Assistant Launches
Google has launched NotebookLM, an AI-powered note-taking application that provides users with a personalized virtual research assistant. NotebookLM instantly becomes an expert on a user's projects by analyzing documents they upload. It aims to help users go from information to insight more quickly by enabling seamless transitions between reading, asking questions, and writing. Users maintain control and privacy over sensitive information. The launch has been met with excitement about the potential for this type of AI application, though some note it is still in the early stages. At the moment, it's only available in US.
Google Launches AI Studio for Easy App and Chatbot Development
Google launched AI Studio, a new web-based tool that allows developers to easily create prompts and chatbots powered by Google's new Gemini AI models. According to VP Josh Woodward, AI Studio aims to be "the fastest way to build with Gemini" and invites developers to "come play with it" to build apps and experiences. The tool features generous free access to Gemini, though Google will review input/output data to improve quality. The launch comes right after the announcement of Gemini last week and its integration into Google's Bard chatbot.
Google Unveils Imagen 2, New AI Image Generator
Google announced the launch of Imagen 2, the second generation of its AI text-to-image model. Imagen 2 introduces new capabilities like rendering text and logos within images, with support for multiple languages including Chinese, Hindi, and Portuguese. The upgraded model also aims to produce higher quality and more detailed images compared to the original Imagen. While specific details on Imagen 2's training data remain undisclosed, Google touts improved safeguards like invisible watermarks to identify AI-generated images.
DeepMind's FunSearch Makes Mathematical Discoveries Using AI
DeepMind researchers introduced FunSearch, a new AI method that can make verifiable discoveries in mathematics and computer science. FunSearch pairs a large language model with an evaluator to iterate towards creative solutions expressed as computer code. For the first time, this approach yielded new solutions for open problems like the "cap set problem" in math. The authors highlight that FunSearch favoring simple, interpretable code allows further insights by scientists. Beyond advancing theory, FunSearch also found better algorithms for practical challenges like bin packing.
Mistral AI Launches Early Access to AI Platform
AI startup Mistral AI opened early beta access to its platform for deploying customizable, generative AI models. The platform currently serves endpoints for text generation and embedding in multiple languages. Mistral AI says its services use strong alignment techniques and optimized models to create pleasant, easy-to-control chatbots.
Microsoft Unveils Phi-2, a Laptop-Friendly AI Model That Rivals Larger Systems
Microsoft Research announced the release of Phi-2, a small generative AI model with just 2.7 billion parameters that nonetheless achieves performance comparable to much larger models. According to Microsoft, Phi-2 outperforms Meta's 7-billion parameter Llama 2-7B and the 7-billion parameter Mistral-7B on certain benchmarks, while delivering higher quality and less toxic responses. However, Phi-2 is currently only available for non-commercial research purposes under a restrictive license.
Atlassian Launches AI Capabilities Across its Platform
Atlassian announced the general availability of its new AI capabilities across its products including Jira, Confluence, and Jira Service Management. Atlassian states that AI will help boost individual productivity by generating content and summaries, automating tasks through natural language, and providing context-specific help.
Create Original Songs with Microsoft Copilot and Suno
Microsoft has partnered with AI music startup Suno to bring song creation capabilities to Copilot. Users can now generate complete, customized songs simply by prompting Copilot with a sentence. The songs include lyrics, instrumentals, and vocals.
AI Masters Complex Physical Game Labyrinth in Just 6 Hours
In an impressive demonstration of AI's capabilities, researchers at ETH Zurich developed an AI robot named CyberRunner that mastered the incredibly difficult physical game Labyrinth in just 6 hours. Labyrinth, which involves navigating a metal ball through a maze by tilting the game board, is known for requiring advanced motor skills and real-time problem solving. By using machine learning algorithms, CyberRunner was able to complete the maze in a record time of 14.8 seconds on December 20, 2023. This accomplishment shows how AI can quickly learn to solve complex physical tasks based on vision, physical interaction, and training.
Interesting Articles and Resources
The article explains in detail how the Gemini presentation video was made.
OpenAI guide to write better prompts and be more effective with artificial intelligence tools.
That wraps up this week's AI news! If you found this informative, don't forget to share it with friends and colleagues. And be sure to subscribe to get next week's news straight to your inbox. Thanks for reading!