Hire Ava, the Industry-Leading AI BDR

Ava automates your entire outbound demand generation so you can get leads delivered to your inbox on autopilot. She operates within the Artisan platform, which consolidates every tool you need for outbound:

• 300M+ High-Quality B2B Prospects

• Automated Lead Enrichment With 10+ Data Sources Included

• Full Email Deliverability Management

• Personalization Waterfall using LinkedIn, Twitter, Web Scraping & More

Book a demo to see what Ava can do.

Today:

• Mira Launches New AI Venture

• Runway Expands Video Editing Capabilities

• Chinese Researchers Unveil LLaVA-o1

• Apple Enhances Siri with AI

• YouTube Launches Dream Screen Feature

• Google Cloud Unveils AI Agents

Ex-OpenAI CTO Murati’s New Strartup is Revealed!

Mira, the former Chief Technology Officer of OpenAI, recently left the company amid internal conflicts and competition from other AI firms. She is launching a new AI venture with top experts, focusing on refining AI models for specific tasks instead of just making them bigger. 

This shift is important as the initial training of AI models is becoming less effective, making specialized improvements crucial. Mira’s team aims to enhance applications in areas like healthcare and education. 

WATCH THE VIDEO ON YOUTUBE

Runway Expands Video Editing Capabilities

Runway, a New York-based AI startup celebrating six years, launched its new ‘Expand Video’ feature. This tool lets users lengthen video clips while keeping visuals smooth, soon available to all through their Gen-3 Alpha Turbo suite. CEO Cristóbal Valenzuela highlighted the feature’s creative potential, allowing effects like zooms and reveals. 

Runway also introduced Act-One for better character performances and partnered with Lionsgate for exclusive AI tools. Their technology was used in the Oscar-winning film "Everything Everywhere All at Once," making special effects easier and cheaper.

Chinese Researchers Unveil LLaVA-o1

Chinese researchers have created LLaVA-o1, a new open-source AI model aiming to compete with OpenAI’s o1. Unlike earlier models that answer questions without clear reasoning, LLaVA-o1 breaks down problem-solving into four steps: summarizing the question, describing relevant image parts, logical reasoning, and concluding the answer. This structured approach helps the AI think more accurately and handle complex tasks better. 

Additionally, LLaVA-o1 uses a method called stage-level beam search to improve response quality by evaluating multiple options at each step. Trained on 100,000 image-question pairs, LLaVA-o1 outperforms other models, setting a new standard for AI reasoning.

Apple Enhances Siri with AI

Apple is upgrading its Siri digital assistant by incorporating advanced large language models (LLMs) to enable more natural and interactive conversations. This enhanced Siri, internally called 'LLM Siri,' aims to compete with OpenAI’s ChatGPT and other voice services by handling more complex requests quickly and efficiently. 

According to sources, the new Siri will support back-and-forth dialogues, making interactions smoother and more intuitive. Although Apple has not officially announced the project, the revamped Siri is expected to be unveiled next year and become available to users in 2026.

YouTube Launches Dream Screen Feature

YouTube is testing a new feature called Dream Screen for creating Shorts, which are short videos on the platform. Dream Screen uses artificial intelligence (AI) to generate unique images or videos based on what you describe. These creations can be used as green screen backgrounds in your Shorts, allowing for more creative and imaginative content. 

Currently, only a limited number of creators in certain countries can access this feature as YouTube refines it. All content made with Dream Screen must follow YouTube’s Community Guidelines to prevent inappropriate or harmful material. Creators should carefully review AI-generated content before publishing to ensure it meets the rules.

Google Cloud Unveils AI Agents

Google Cloud has launched AI Agent Space, a new platform that allows businesses to create, deploy, and collaborate on AI agents to automate tasks and enhance customer experiences. This initiative places Google alongside competitors like Microsoft, SAP, and Salesforce in the growing AI market. AI Agent Space provides partners with tools, resources, and support to develop customized AI agents, which are then promoted through Google Cloud Marketplace to reach more users. 

Currently, 19 agents from partners such as Accenture and Deloitte are available, with plans to add hundreds more. Unlike rivals, Google emphasizes flexibility and an open ecosystem, aiming to foster innovation and meet diverse business needs.

🧠RESEARCH

Enhancing the Reasoning Ability of Multimodal Large Language Models via Mixed Preference Optimization

Researchers boosted multimodal language models’ reasoning by developing a large-quality dataset and a method called Mixed Preference Optimization. This helps models understand and use both text and images better. Their improved model scores higher on tests, matching much bigger models. They will make their tools publicly available.

Multimodal Autoregressive Pre-training of Large Vision Encoders

AIMV2, a vision encoder combining images and text for multimodal pre-training. It uses a simple, scalable method to generate raw image patches and text tokens. AIMV2 achieves top performance in vision tasks like classification and localization, surpassing leading models like CLIP. Its versatility enhances multimodal understanding significantly.

Marco-o1: Towards Open Reasoning Models for Open-Ended Solutions

Marco-o1 advances reasoning models by tackling open-ended problems beyond standard disciplines like math and coding. Using Chain-of-Thought fine-tuning, Monte Carlo Tree Search, and reflection mechanisms, it explores solutions where standards are unclear and rewards are hard to measure. Marco-o1 aims to excel in complex, real-world problem-solving.

Hymba: A Hybrid-head Architecture for Small Language Models

Hymba introduces a hybrid-head architecture for small language models, combining transformer attention for detail recall with state space models for efficient context summarization. Using learnable meta tokens and optimized mechanisms like cross-layer KV sharing, it achieves superior performance with reduced cache size and higher efficiency, outperforming larger models in benchmarks.

OpenScholar: Synthesizing Scientific Literature with Retrieval-augmented LMs

OpenScholar is a retrieval-augmented language model designed to assist researchers by synthesizing citation-backed answers from 45 million open-access papers. It outperforms GPT-4o and PaperQA2 in accuracy and citation reliability. With tools like ScholarQABench and self-feedback loops, it enhances literature searches, earning expert preference over human-written responses in evaluations.

🛠️TOP TOOLS

Rewind - Browse, search, and ask Rewind about anything you’ve seen on your phone. Rewind is a truly personalized AI in your pocket.

Lovable - Superhuman full stack engineer.

Wondershare Virbo - AI-Powered One-Stop Video Marketing Solution to Maximize Your Conversions

HumanLayer - API and SDK that enables AI Agents to contact humans for feedback, input, and approvals.

RemoteBase - Quality-assured AI trainers, pre-vetted by experts to elevate your AI’s potential with speed and skill.

📲SOCIAL MEDIA

🗞️MORE NEWS

• Lightricks launched LTXV, an open-source AI model for generating high-quality videos in seconds, aiming to challenge tech giants by promoting innovation through accessibility, speed, and collaboration, targeting creators and smaller studios globally.

• Rabbit’s R1 device now features "teach mode," enabling users to train its AI to perform custom tasks by demonstrating them, enhancing automation across platforms like Spotify and YouTube. However, CAPTCHA-related limitations persist.

• Wordware raised $30 million to simplify AI development by enabling users to create AI agents using natural language. The platform targets enterprises and individuals, aiming to revolutionize AI with faster, accessible solutions.

• MIT researchers developed an efficient algorithm for training AI systems to handle complex, variable tasks. By focusing on key tasks, their approach improves reliability, reduces training costs, and enhances performance across applications like traffic management.

• Microsoft's Recall AI feature for Copilot Plus PCs allows users to search activities using snapshots and natural queries. Available to Windows Insiders, it prioritizes security with opt-in functionality, encrypted data, and no cloud sharing.