What's New with AI #3-2025
Hello AI Aficionados!
Welcome to the latest edition of AI Horizon, your go-to source for the most recent developments in artificial intelligence. In this issue, we cover groundbreaking releases, strategic initiatives, and innovative projects shaping the AI landscape. Stay informed with concise summaries of the top stories from the past week.
1. UC Berkeley's NovaSky Team Launches Sky-T1-32B-Preview
OThe NovaSky team at UC Berkeley has released Sky-T1-32B-Preview, an open-source reasoning model that rivals earlier versions of OpenAI’s o1.
Key Details:
Model Foundation: Fine-tuned from Alibaba’s Qwen2.5-32-Instruct.
Training Data: Generated using the open-source reasoning model QwQ-32B-Preview.
Training Efficiency: Completed in 19 hours on 8 H100 GPUs, costing approximately $450.
Performance: Matches or exceeds earlier versions of OpenAI's o1, especially in mathematics and coding tasks.
Open-Source Commitment: Entire pipeline, including training data, code, and model weights, is fully open source.
2. OpenAI Ventures into Robotics Hardware Development
OpenAI has posted its first robotics hardware job listings, indicating plans to develop custom robots with advanced AI capabilities.
Key Details:
Leadership: Caitlin Kalinowski, former Meta AR glasses lead, is heading the initiative as OpenAI's hardware director since November.
Recruitment: Hiring for roles in sensor suite development, mechanical design, and lab operations management.
Objectives: Developing 'general-purpose robots' for dynamic real-world settings, exploring various robotic form factors.
Background: OpenAI previously disbanded its robotics team in 2020, which had worked on projects like training a robotic hand to solve a Rubik’s Cube.
3. OpenAI Introduces 'Tasks' Feature in ChatGPT
4. Microsoft Enhances AutoGen Platform with v4 Updates
Microsoft has announced significant updates to its AutoGen platform, facilitating more efficient development of AI agents.
Key Details:
Asynchronous Messaging: Enables bots to communicate across different coding languages effectively.
Improved Transparency: Provides developers with better insights into agent decision-making processes, simplifying debugging and updates.
Expanded Capabilities: Bots can now perform tasks across the entire workspace, with enhanced memory for extended planning.
5. François Chollet Launches Ndea to Pursue AGI
François Chollet, creator of the Keras AI framework, has introduced Ndea, a new AI lab aiming to achieve Artificial General Intelligence (AGI) through an alternative research approach, alongside Zapier founder Mike Knoop.
Key Details:
Research Strategy: Combines deep learning with program synthesis to develop AI that learns and adapts with human-level efficiency.
Positioning: Offers an alternative to large-scale deep learning approaches, addressing current AI limitations due to training data constraints.
Goals: Plans to build a "factory for rapid scientific advancement," focusing on areas like drug discovery and unexplored territories.
Related Initiatives: Chollet recently launched the ARC Prize Foundation, a nonprofit developing benchmarks to evaluate human-level AI capabilities.
6. Microsoft Introduces Microsoft 365 Copilot Chat
Microsoft has unveiled Microsoft 365 Copilot Chat, a rebranded free entry-level version of its AI assistant platform, introducing pay-per-use agent capabilities while maintaining core AI features for business users.
Key Details:
Features: Offers free access to GPT-4o-powered chat, including web-based knowledge, file analysis capabilities, and image and code generation.
Custom AI Agents: Provides access to custom AI agents for task automation, with a consumption-based model at $0.01 per message or $200 for 25,000 messages monthly.
Integration: Agents can leverage knowledge sources for various tasks, with the Copilot Control System allowing easy platform management by IT teams.
Objective: Bridges the gap between free users and the full Microsoft 365 Copilot subscription, priced at $30 per user per month.
7. Luma Labs Releases Ray 2 AI Video Model
Luma Labs has released Ray 2, the startup’s next-generation AI video model, offering unprecedented motion quality and physics realism through a new multimodal architecture trained with 10 times more computing power than its predecessor.
Key Details:
Capabilities: Generates high-quality video clips up to 10 seconds long from text prompts, demonstrating advanced motion and physics understanding.
Sophistication: Exhibits a sophisticated grasp of object interactions, from natural scenes like water physics to complex human movements.
Functionality: Currently handles text, image, and video-to-video generations, with editing capabilities to be added soon.
Availability: Launching first in Luma’s Dream
8. NVIDIA CEO on AI Chips
Jensen Huang, CEO of NVIDIA, has revealed that AI chip development is now surpassing Moore's Law. This suggests a rapid evolution in AI hardware, promising groundbreaking innovations in the near future.
AI Models and Upgrades
Google’s Gemini AI: Google continues to refine its Gemini models, making them smarter and more cost-effective. In addition, Google DeepMind’s AlphaChip is pushing the boundaries of chip design.
Meta’s Llama 3.2: Meta has introduced Llama 3.2, a vision-capable AI model, demonstrating significant strides in multimodal AI capabilities.
Google TV Enhancements: Google TV now integrates Gemini AI, enhancing voice control and search for a seamless user experience.
On the House
Unlocking the Power of AI: Top 17 Frameworks and Libraries for Large Language Models
Why Less Can Be More in AI: Understanding SLMs (Small Language Model)
That's all for now! Stay tuned for more exciting updates next week. Keep innovating, and remember, the future is now! 🌟👋
Stay curious, stay inspired,
Together, we're not just witnessing the future; we're creating it. Stay tuned for more insights and stories in our next edition! 🌟🛠️