Good morning. It’s Friday, September 20th.

Did you know: 35 years ago today, Apple Computer introduced the Macintosh Portable?

In today’s email:

• YouTube Shorts Integrates Google's Veo AI

• Nvidia's Jim Fan Predicts Robotics Breakthrough

• KLING AI Video Generator Releases Version 1.5

• OpenAI's o1 Models Top Chatbot Arena

• 5 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

YouTube Shorts to integrate Veo, Google’s AI video model

At its Made On YouTube event, the platform announced the integration of Google DeepMind’s Veo into Shorts. This addition enables creators to produce high-quality backgrounds and six-second video clips. Veo, introduced at Google I/O 2024, competes with other video generation models like OpenAI’s Sora and supports 1080p video creation in various styles, with features for editing and remixing footage.

This integration improves upon the existing “Dream Screen” feature, which previously allowed background generation from text prompts. With Veo, creators can now also generate standalone six-second clips, enhancing content with smoother transitions and contextually relevant scenes. The feature will include watermarked AI-generated content using DeepMind’s SynthID technology. Read more.

Nvidia researcher Jim Fan expects "GPT-3 moment" for robotics in the next few years

Jim Fan, senior researcher at Nvidia, anticipates significant progress in robotics over the next two to three years, envisioning a "GPT-3 moment" for foundational models in the field. Fan's team is developing foundational AI models for humanoid robots through Project Groot, emphasizing their potential in environments designed for human interaction. However, Fan stresses that technical progress alone isn’t enough—affordability, mass production, safety, and regulatory frameworks constitute formidable barriers to widespread adoption.

Nvidia's strategy combines internet data, simulated environments, and real-world robot data to cultivate more adaptable AI models. The team employs "Eureka" to automate robot training, which could enable a single AI model to control both virtual and physical agents. While Fan expresses optimism, he notes that refining data pipelines and integrating unconscious motor skills with conscious planning are crucial to fully realizing the potential of these advancements. Read more.

AI video generator KLING releases version 1.5 with impressive features

Kuaishou has released version 1.5 of its AI video generator KLING, bringing higher resolution and new features. Users can now create videos in 1080p HD in professional mode, with the new version offering a 95% performance improvement over its predecessor. Enhancements include sharper visuals, improved dynamics, and more relevant prompt handling, showcased in sample videos with realistic lighting and detailed textures, such as lifelike fur.

A standout addition is the "Motion Brush" feature, allowing users to precisely control the movement of individual video elements. KLING 1.5 also supports multiple aspect ratios, including landscape, portrait, and square formats, with Motion Brush videos capped at five seconds.

First launched globally in 2023, KLING continues to offer 66 free video creation credits daily, with no changes to pricing in this latest update. Read more.

Chatbot Arena: OpenAI o1-preview and o1-mini dominate the competition

OpenAI's recent AI models, o1-preview and o1-mini, have excelled in the Chatbot Arena, securing top rankings in overall performance, safety, and technical capabilities. o1-preview claimed first place across all evaluated categories, while o1-mini, optimized for STEM tasks, shared second place with the GPT-4o model. The evaluation, based on over 6,000 community ratings, highlighted the models' strengths in mathematics, complex prompts, and programming.

However, the o1 models garnered significantly fewer ratings than established competitors like GPT-4o and Claude 3.5, each receiving just under 3,000 reviews. This limited sample size raises concerns about the validity of the results and potential biases in the evaluations. Although o1 is designed to enhance AI reasoning by allowing longer processing times, it does not consistently outperform GPT-4o, especially in tasks that require rapid responses. Read more.

AI Model Releases and Open-Source Initiatives

• China's Alibaba launches over 100 new open-source AI models, releases text-to-video generation tool

• Try out OpenAI o1 in GitHub Copilot and Models

• Kyutai Open Sources Moshi—A Game-Changing Full-Duplex Dialogue System for Real-Time Conversations

• French AI startup Mistral overhauls its chat service

• Microsoft’s GRIN-MoE AI model takes on coding and math, beating competitors in key benchmarks  

• Anthropic Introduces Contextual Retrieval

AI Governance, Ethics, and Policy

• US to host global AI safety summit in San Francisco

• UN advisory body makes seven recommendations for governing AI

• LinkedIn scraped user data for training before updating its terms of service

• "AGI system could be built in as little as three years": Ex-OpenAI employee warns US Senate

Entertainment and Media

• Lionsgate signs deal to train AI model on its movies and shows

• Tencent's EzAudio AI transforms text to lifelike sound, sparking innovation and debate

• You can now exclude music styles when generating with Suno

Wut

• Quantum computers teleport and store energy harvested from empty space

📄 Qwen2.5-Coder Technical Report

📄 InfiMM-WebMath-40B: Advancing Multimodal Pre-Training for Enhanced Mathematical Reasoning ​​

📄 LLMs + Persona-Plug = Personalized LLMs

📄 Fine-Tuning Image-Conditional Diffusion Models is Easier than You Think  

📄 A Controlled Study on Long Context Extension and Generalization in LLMs

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.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email!