How can we make AI action generation both faster and more expressive without compromise? Researchers from Tsinghua University, Berkeley AI Research (BAIR), and The University of Hong Kong unveil their new Mean Velocity Policy (MVP). This innovative method models the "mean
RESEARCH
-
PaddleOCR Surpasses Tesseract as Most-Starred OCR Project
By
–
PaddleOCR has surpassed 73.3K stars on GitHub—overtaking Google’s Tesseract (73.2K) to become the most-starred OCR project globally. GitHub: https://
github.com/PaddlePaddle/P
addleOCR
… PP-OCRv5: https://
arxiv.org/pdf/2603.24373 PaddleOCR-VL: https://
arxiv.org/pdf/2603.24326 -
Six Types of AI Models in Machine Learning
By
–
6 Types of #AI Models
by @PythonPr #ArtificialIntelligence #MachineLearning #ML -
AI will make expert knowledge accessible to all
By
–
I would expect that a lot of things that were old hat to experts, but completely inaccessible to most people, will go viral in the coming months. Sure, anyone could have done those things before, but it required a lot of deep knowledge. Now the AI can make it happen by asking.
-
Hark: New AI Lab Building Proactive Multimodal Personal AI
By
–
Meet Hark – the most ambitious AI lab you haven't heard of yet. Founded by serial entrepreneur Brett Adcock, Hark is building AI that's proactive, personalized, and speaks through voice, text, vision & memory. 45+ researchers from Apple, Meta, Google & Tesla iPhone's
-
ByteDance Releases DeerFlow 2.0 Open-Source SuperAgent on GitHub
By
–
ByteDance dropped an open-source SuperAgent — and it hit #1 on GitHub Trending. DeerFlow 2.0 can: Deep web research with cited sources Generate full reports with charts, images & video Run Python & bash in a secure sandbox Create slide decks & UI components
-
Luma Launches Uni-1 Multimodal Image Model Beating Google and OpenAI
By
–
Luma just launched a multimodal image model that beats Google and OpenAI on benchmarks. Uni-1 is built on a decoder-only transformer trained on images, video, audio, language AND spatial reasoning. #1 in human preference Elo for Style & Editing 10–30% cheaper than
-
OPUS: Intelligent Data Selection for LLM Pre-training
By
–
Is there a smarter way to pick data for training Large Language Models? Researchers from multiple institutions, led by Shaobo Wang, introduce OPUS. This novel method dynamically and intelligently selects the most impactful data for LLM pre-training in every single training
-
Natural-Language Agent Harnesses Improve Model Performance Architecture
By
–
“Natural-Language Agent Harnesses” This paper argues that agent performance increasingly depends on the harness around the model, but that harness logic is usually buried in controller code and runtime-specific conventions. So they propose Natural-Language Agent Harnesses
-
Robots Learn Through Training, Not Programming Anymore
By
–
𝗥𝗼𝗯𝗼𝘁𝘀 𝗮𝗿𝗲 𝗻𝗼 𝗹𝗼𝗻𝗴𝗲𝗿 𝗽𝗿𝗼𝗴𝗿𝗮𝗺𝗺𝗲𝗱.
— Pascal Bornet (@pascal_bornet) 30 mars 2026
𝗧𝗵𝗲𝘆 𝗮𝗿𝗲 𝘁𝗿𝗮𝗶𝗻𝗲𝗱.
That’s the real shift I’m seeing from NVIDIA’s GTC.
Using Isaac Lab, robots are learning through reinforcement learning in simulation:
▪️ Millions of trials
▪️ No step-by-step… pic.twitter.com/JcrIMGW0Lw𝗥𝗼𝗯𝗼𝘁𝘀 𝗮𝗿𝗲 𝗻𝗼 𝗹𝗼𝗻𝗴𝗲𝗿 𝗽𝗿𝗼𝗴𝗿𝗮𝗺𝗺𝗲𝗱. 𝗧𝗵𝗲𝘆 𝗮𝗿𝗲 𝘁𝗿𝗮𝗶𝗻𝗲𝗱. That’s the real shift I’m seeing from NVIDIA’s GTC. Using Isaac Lab, robots are learning through reinforcement learning in simulation: ▪️ Millions of trials ▪️ No step-by-step instructions ▪️ Learning by reward and feedback That’s how a machine learns to drive, jump, flip… and recover. What stands out to me is this: 𝗪𝗲’𝗿𝗲 𝗺𝗼𝘃𝗶𝗻𝗴 𝗳𝗿𝗼𝗺 𝗰𝗼𝗱𝗲 → 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴. And once systems can learn, they don’t just execute. They adapt. Same shift we saw with LLMs. Now it’s happening in the physical world. 𝗦𝗼 𝗵𝗲𝗿𝗲’𝘀 𝗺𝘆 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻: Are we ready for machines that improve faster than we can program them? #ai #robotics #reinforcementlearning #nvidia #gtc #futureofwork
→ View original post on X — @pascal_bornet, 2026-03-30 05:01 UTC