What if robots could improve themselves by learning from their own failures in the real-world? Introducing 𝗣𝗟𝗗 (𝗣𝗿𝗼𝗯𝗲, 𝗟𝗲𝗮𝗿𝗻, 𝗗𝗶𝘀𝘁𝗶𝗹𝗹) — a recipe that enables Vision-Language-Action (VLA) models to self-improve for high-precision manipulation tasks. PLD couples real-world residual reinforcement learning with standard supervised fine-tuning — letting robots discover, recover, and distill their own data flywheel. Quick 🧵

Zero teleoperation. Zero real-world data. ➔ Autonomous humanoid loco-manipulation in reality. Introducing VIRAL: Visual Sim-to-Real at Scale. We achieved 54 autonomous cycles (walk, stand, place, pick, turn) using a simple recipe: 1. RL 2. Simulation 3. GPUs Website: https://t.co/tmkKpWsqqW Arxiv: https://t.co/PKIfxToUGw Deep dive with me: 🧵

Reality of robotics: humanoid kung fu is solved before they can open doors with RGB. Here we are. Introducing the frontier of sim2real at NVIDIA GEAR. 100% sim data. RGB input only. Code name: 𝗗𝗼𝗼𝗿𝗠𝗮𝗻. We are opening the sim-to-real door. https://t.co/ar1dREHsRi 🧵 https://t.co/4d96Fl9qqv

You’re not wrong. https://t.co/c0trHwARee

controversial recognition of a “central” New Jersey that also functions as an acknowledgement that it’s not really a separate region https://t.co/CvxWM6lTXf

Steve Buscemi dressed as John Waters for John Waters’ Xmas card https://t.co/PPPFlxww03

Steve Buscemi dressed as John Waters for John Waters’ Xmas card https://t.co/PPPFlxww03

https://t.co/DuQvsnBGkC

https://t.co/DuQvsnBGkC

Is the West okay? https://t.co/4qR88wNEjd

ELON: “VIRTUE SIGNALING KILLS” Elon gets serious when asked why he called Sen. Mark Kelly a "Traitor" for pushing to send more U.S. aid and weapons to Ukraine. He’s calling out what he sees as moral theater around the war, arguing that thousands keep dying in trenches with “no movement in the lines” and no realistic path to stopping it. It’s an uncomfortable take, but it cuts through slogans and forces the question many avoid: if there’s no plan to end it, what exactly are all these lives being lost for? “We should have empathy for the thousands of people dying everyday in the trenches, for no movement in the lines, for the past two years thousands of people have died every week for nothing. I take great offense at those who put the appearance of goodness over the reality of it. Those who virtue signal and say we can't give into Russia, but have no solution to stopping thousands of kids dying every day. I have contempt for such people and I want to make that clear, because they're virtue signaling and their lack of a solution means that kids don't have a father… It means parents lost a son. For what? Nothing." Source: @elonmusk, Fox News

Grok models have been dominating the #1 spot on the OpenRouter and Kilo Code leaderboards since taking the top position four months ago Processing trillions of tokens every month, with a multi-trillion-token usage gap over competitors And it is not even close https://t.co/PyRdiIBsc0

Friend of mine shared this - "I finally convinced my mom to buy a @Tesla after working there for 5 years. She did not want FSD but I made her get it anyway. After a few months of ownership, she gained the courage to try it and this was the result!" https://t.co/gvuGruSNe0

BREAKING: Grokipedia just launched a new feature that lets anyone suggest new articles directly on the site. Users can now submit topics, explain why they matter, and shape what Grokipedia covers next. https://t.co/F9DDvDAN0Z

You can now suggest new articles on Grokipedia if you cannot find one Just select “Suggest Article”, enter the topic you want covered, and submit Your suggestions help expand Grokipedia....Contribute easily https://t.co/PIlxbiXjrx

Quantum superposition isn’t just science; it’s a bridge to the spiritual. 🌌 In quantum mechanics, a particle can exist in multiple states simultaneously until observed. This mirrors the spiritual idea that reality is fluid, interconnected, and shaped by consciousness. Just as superposition collapses into a definite state upon observation, our thoughts and awareness might shape the universe around us. 🤔 Credit; Illusion_durality #QuantumMechanics #Spirituality #Consciousness

This paper is a big deal! It's well known that RL works great for math and code. But RL for training agents is a different story. The default approach to training LLM agents today is based on methods like ReAct-style reasoning loops, human-designed workflows, and fixed tool-calling patterns. The issue is that these methods treat the environment as passive rather than interactive. But in the real world, agents must make sequential decisions, maintain memory across turns, and adapt to stochastic environmental feedback. That's fundamentally an RL problem. This new research introduces Agent-R1, a framework for training LLM agents with end-to-end reinforcement learning across multi-turn interactions. As agents move from predefined workflows to autonomous interaction, end-to-end RL becomes the natural training paradigm. Agent-R1 provides a modular foundation for scaling RL to complex, tool-using LLM agents. Standard RL for LLMs assumes deterministic state transitions. You generate a token, append it to the sequence, done. But agents trigger external tools with uncertain outcomes. The environment responds unpredictably. State transitions become stochastic. Therefore, the researchers extend the Markov Decision Process framework to capture this. State space expands to include full interaction history and environmental feedback. Actions can trigger external tools, not just generate text. Rewards become dense, with process rewards for intermediate steps alongside final outcome rewards. Two core mechanisms make this work. An Action Mask distinguishes agent-generated tokens from environmental feedback, ensuring credit assignment targets only the agent's actual decisions. A ToolEnv module manages the interaction loop, handling state transitions and reward calculation when tools are invoked. On multi-hop question answering, RL-trained agents dramatically outperform baselines. The weakest RL algorithm (REINFORCE++) still beat naive RAG by 2.5x on average EM. GRPO achieved 0.3877 average EM compared to 0.1328 for RAG. Ablation results also confirm that the design matters. Disabling the advantage mask dropped PPO performance from 0.3719 to 0.3136. Disabling the loss mask caused further degradation to 0.3022. Precise credit assignment is essential for multi-turn learning. Paper: https://t.co/BrIBT3AAxC Learn to build effective AI agents in my academy: https://t.co/JBU5beIoD0

How AI Can Accelerate Systems Performance Research This new research introduces AI-Driven Research for Systems (ADRS), a framework where LLMs iteratively generate, evaluate, and refine algorithms for systems performance problems automatically. The researchers applied three open-source ADRS frameworks (OpenEvolve, GEPA, ShinkaEvolve) across ten real research tasks spanning networking, databases, and distributed systems. What did they find? In MoE load balancing, ADRS discovered an algorithm that's 13x faster than the best-known proprietary implementation. In multi-region cloud scheduling with spot instances, it achieved 35% greater cost savings than an expert-developed baseline. In transaction scheduling, it improved the makespan by 60% over state-of-the-art for the offline case. ADRS borrowed Hamilton's Apportionment method from political science to solve GPU load balancing. It applied Borda Count from voting theory to optimize transaction ordering. It used Kirchhoff's Current Law from electrical engineering to repair network telemetry. The cost? Most tasks completed in under 5 hours for less than $30. Systems problems are uniquely suited for AI-driven research because candidate solutions can be verified automatically. The LLM proposes, the system evaluates. There is no human judgment bottleneck. The researchers outline best practices across three axes: - For specifications: less is more, and more is less. Structured prompts with clear problem definitions, evaluation criteria, and context. - For evaluation: diverse test sets and precise scoring functions prevent reward hacking. - For feedback: calibrated granularity provides actionable guidance without overfitting. As AI takes on algorithm discovery, researcher effort shifts from solution design to problem formulation and strategic oversight. The 40% of research time spent on solution iteration can now be automated, fundamentally changing how systems research is done. Paper: https://t.co/YKcVKuku4P Learn to build effective AI agents in our academy: https://t.co/zQXQt0PMbG

You sent a letter to @elonmusk threatening him with legal action if he dare host an interview with a candidate for President of the United States. How dare you engage in such belligerence against our democratic elections and then claim *you* are now the victim. FAFO. https://t.co/Vc5gB2T7dZ

한국 테슬라 오너들이 단 1개월 만에 FSD(감독형)으로 누적 주행거리 100만 km를 돌파했습니다. 이는 대한민국을 약 480바퀴 돌고도 남는 거리입니다! *섬을 제외한 공식 해안선 길이와 북쪽 국경 기준, 1바퀴당 약 2,413 km Korea Tesla owners have surpassed 1 million km of cumulative driving distance with FSD (Supervised) in just one month. This distance is enough to circle the entire country of South Korea approximately 480 times—with some to spare! *Based on the official coastline length (excluding islands) and the northern border, one full lap is approximately 2,413 km.

Maniac who gouged a woman’s eye out in Seattle has a long criminal history - 8 arrests 2025 - 1 arrest 2024 - 4 arrests 2023 - 1 arrest 2020 - stabbing 2011 Including for assault, indecent exposure, drug offenses, property destruction, and weapons charges JAIL THE JUDGES https://t.co/E8tzAJe3io

2026: Mad Max all the way to solving autonomous driving for good. Humanity: zero takeover asymptotically. https://t.co/Bd1t9J0CoH

NextStep-1.1 is out https://t.co/OO4dTB5yLc https://t.co/XNCkf1cf9T

+1 more

The growth story of Boston Dynamics' Atlas →Last Christmas, it did a backflip. 🤶 →This Christmas, it's raising a toast to you! 🍻 Atlas is no longer pursuing high-dynamic performances.Its goal is to solve specific problems in the real world through cognitive and intelligent actions. BTW,you can meet it at CES 2026.

Every Native born into this world is a victory against colonialism & attempted genocide. You are the resistance. You are hope made flesh. https://t.co/rcdEdU7Bd5

If You're a huge fan of Native Culture can I get a big YESS !!!!!💖💖💖 https://t.co/Ix9G9GOnZO

If You’re a huge fan of Native Culture can I get a big YESS !!!!!💖💖💖 https://t.co/AHB1cf1fUK

Everyone is sleeping on MiniMax's new LLM! Devs are calling it "Claude at 10% the cost" - 72.5% SWE-Multilingual. Beats Sonnet 4.5 - 88.6% VIBE-bench. Beats Gemini 3 Pro I used it to build a stock analyst that generates code, executes it & returns insights. 100% open-source! https://t.co/nRmmwUYtJX

How does an embryo reliably "compute" its form - "cell by cell" - using only local interactions and mechanics, yet produce a precise global body plan? I’m excited to share our Nature Methods paper "MultiCell: geometric learning in multicellular development", presenting #AIxBiology research led by @HaiqianYang and the result of a great collaboration with Ming Guo, George Roy, Tomer Stern, Anh Nguyen and Dapeng Bi. A long-standing challenge in developmental biology is to predict how thousands of cells collectively self-organize as tissues fold, divide, and rearrange. In MultiCell, we represent a developing embryo as a dual graph that unifies two complementary views of tissue mechanics with single-cell resolution: cells as moving points (granular) and cells as a connected foam (junction network). This lets the model learn dynamics from both geometry and cell–cell connectivity. On whole-embryo 4D light-sheet movies of Drosophila gastrulation (~5,000 cells), our model predicts key cell behaviors and the timing of events, including junction loss, rearrangements, and divisions with high accuracy, at single-cell resolution. Beyond prediction, the same representation supports robust time alignment across embryos and offers interpretable activation maps that highlight the morphogenetic "drivers" of development. The broader goal is a foundation for cell-by-cell forecasting in more complex tissues, and eventually for detecting subtle dynamical signatures of disease. Kudos to the team for this inspiring collaboration with brilliant researchers to push the boundary of AI for biology! Citation: Yang, H., Roy, G., Nguyen, A.Q., Buehler, M.J., et al. MultiCell: geometric learning in multicellular development. Nature Methods (2025), DOI: 10.1038/s41592-025-02983-x Code/data links are in the manuscript.

@lukas_m_ziegler The prices of everything in technology always start real high and come down real quick. I remember when a color printer was $45,000 in 1989, and today a $79 printer is way, way better. But how does https://t.co/JuDkLaml2o sell Neo for $20,000?

Introducing ✨VideoRAG✨ - Extremely Long Video Understanding. Chat with your videos like never before. Accepted by KDD 2026! 🚀 🔗 Fully Open Source: https://t.co/QrNajpOby4 Current AI struggles with long videos. Most systems can't handle hours of content or understand connections across multiple videos. ✨ VideoRAG breaks these barriers! - Easy-to-Use: Just drag & drop your videos and start chatting - No Length Limits: Handle everything from 30-second clips to 100+ hour documentaries - Cross-Modal Understanding: Understands visual content, audio, and context together --------------------------------------------------------- For Developers & Researchers: - 🚀 Process hundreds of hours on a single RTX 3090 - 🧩 Multi-modal knowledge graphs for structured understanding - 📊 LongerVideos benchmark with 134+ hours across lectures, docs, and entertainment - 🔓 Open source and extensible

@din0s_ @typelessdotcom Here is @blevlabs talking to his agents. And he is a genius. https://t.co/pyxvrissKW