In partnership with

Good morning. It’s Friday, February 7th.

Did you know: On this day in 1984, astronauts Bruce McCandless II and Robert L. Stewart conducted the first untethered spacewalk using the Manned Maneuvering Unit (MMU)

In today’s email:

• Programming Robotics With Natural Language

• Google’s Gemini 2.0 Pro

• ChatGPT Displays Chain-of-Thought

• 4 New AI Tools

• Latest AI Research Papers

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

HuggingFace Pi0 Makes Robotics Fluent in Human Language

Hugging Face and Physical Intelligence just dropped Pi0, an open-source robotics model that cuts out the coding middleman. Robots can now execute complex tasks—folding laundry, bussing tables, packing groceries—straight from natural language commands. Trained across seven robotic platforms and 68 task domains, Pi0 wields flow matching to generate real-time action trajectories at 50Hz, making movements seamless and adaptable.

A faster variant, Pi0-FAST, incorporates frequency-space action sequence tokenization (FAST), accelerating training speeds fivefold. The model's open-source release lowers barriers to robotic automation, allowing enterprises to fine-tune policies for diverse applications. While promising, Pi0 faces hurdles in computational efficiency and execution reliability. Available on HuggingFace. Read more.

Google launches Gemini 2.0 Pro, Flash-Lite and connects reasoning model Flash Thinking to YouTube, Maps and Search

Google has dropped Gemini 2.0 Flash into general availability, pushed the ultra-cost-efficient Flash-Lite into public preview, and rolled out an experimental Gemini 2.0 Pro. Designed for enterprise-scale efficiency, these models supercharge multimodal processing across Google AI Studio and Vertex AI. Flash-Lite, the leanest model yet, outperforms Gemini 1.5 Flash on MMLU Pro (77.6% vs. 67.3%) and Bird SQL (57.4% vs. 45.6%) while holding firm on a 1-million-token context window—four times the industry standard. Gemini 2.0 Flash goes head-to-head with OpenAI’s o3-mini, maximizing efficiency without slashing capability.

Google has also embedded its Gemini 2.0 Flash Thinking model into Maps, YouTube, and Search, injecting AI-driven reasoning straight into its ecosystem. Meanwhile, Gemini 2.0 Pro—armed with a 2-million-token context window—levels up tool use, reasoning, and code execution. Read more.

OpenAI teases deeper transparency while expanding ChatGPT’s memory

OpenAI has updated the Chain-of-Thought (CoT) reasoning in its O3-Mini models, making it more transparent for both free and paid users. The refinement extends to O3-Mini-High for paid users. However, OpenAI has clarified that these are not the raw CoT outputs, suggesting that while the update offers greater visibility, it stops short of full disclosure.

Separately, OpenAI has expanded ChatGPT’s memory limit by 25% for Plus, Pro, and Team users, with plans to roll out the increase to Enterprise and Edu accounts soon. Read more.

• Boston Dynamics joins forces with its former CEO to speed the learning of its Atlas humanoid robot

• Not every AI prompt deserves multiple seconds of thinking: how Meta is teaching models to prioritize

• Amazon to revamp Alexa’s AI core with Titan and Claude, ready for February launch

• Oxford Teleports Logical Gates, Cracking Quantum Scalability

• ChatGPT Search Now Accessible Without Account Sign-In

• THAAD 6.0: US plans hypersonic-killing, AI-driven, laser-armed missile defense system

• Google has ‘very good ideas’ for native ads in Gemini

• Andrej Karpathy drops 3+ Hour LLM comprehensive review

• DeepSeek: The countries and agencies that have banned the AI company's tech

• Drop yourself into any scene with Pikadditions—no green screen required

• OpenAI expands search for AI data center sites across 16 states

• OpenAI co founder John Schulman is joining Mira Murati’s startup after brief stint at Anthropic

• OpenAI Set to Make Super Bowl Ad Debut

• Musk’s DOGE Puts AI in GSA’s Toolbox as Feds Go Digital

• Google have done it, Gold-medalist Performance in Solving Olympiad Geometry with AlphaGeometry2

• Google Expands Generative AI with Imagen 3 in Gemini API

• An economist used o1-pro to generate a paper in an hour, and got it published in an peer-reviewed journal

• Topaz Labs Unveils Project Starlight, First AI Model for Video Restoration

• Copilot expands with self-iterating AI and automated debugging

• Meta’s PARTNR project takes home robotics beyond gimmicks

• World's first quantum large language model launched, can shape future of AI

• Amazon's profit is suddenly going through the roof, fueling CEO Andy Jassy's $100 billion AI spending spree to fight Google and Microsoft

• Reliable ‘reasoning‘ AI agents may be just around the corner thanks to DeepSeek’s innovations, say researchers

• Musk Called Drones, AI the Future of War in West Point Interview

• Meta launches new program to improve speech and translation AI

📄 UltraIF: Advancing Instruction Following from the Wild

📄 TwinMarket: A Scalable Behavioral and SocialSimulation for Financial Markets​​

📄 Demystifying Long Chain-of-Thought Reasoning in LLMs  

📄 Analyze Feature Flow to Enhance Interpretation and Steering in Language Models

📄 LIMO: Less is More for Reasoning

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 or DM us on 𝕏!