“Those who find beauty in all of nature will find themselves at one with the secrets of life itself.”    ~ L. Wolfe Gilbert https://t.co/SaCQIW3O8v

Introducing NitroGen, an open-source foundation model trained to play 1000+ games: RPG, platformer, battle royale, racing, 2D, 3D, you name it! We are on a quest for general-purpose embodied agents that master not only the real world physics, but also all possible physics across a multiverse of simulations. We found that our GR00T N1.5 architecture, originally designed for robotics, can be adapted easily to play lots of games with wildly different mechanics. Our recipe is simple and bitter lesson-pilled: (1) a 40K+ hour, high-quality dataset of public in-the-wild gameplay; (2) a highly capable foundation model for continuous motor control; (3) a Gym API that wraps any game binary to run rollouts. Our data curation is a lot of fun: it turns out that gamers love to show off their skills by overlaying real-time gamepad control on a video stream. So we train a segmentation model to detect and extract those gamepad displays and turn them into expert actions. We then mask out that region to prevent the model from exploiting a shortcut. During training, a variant of GR00T N1.5 learns to map from 40K hours of pixels to actions through diffusion transformers. NitroGen is only the beginning, and there's a long way to hill-climb on the capability. We intentionally focus only on the System 1 side: the "gamer instinct" of fast motor control. We open-source *everything* for you to tinker: pretrained model weights, the entire action dataset, code, and a whitepaper with solid details. Today, robotics is a superset of hard AI problems. Tomorrow, it might become a subset, a dot in the much larger latent space of embodied AGI. Then you just prompt and "ask for" a robot controller. That might be the end game (pun intended). NitroGen is co-led by our brilliant minds: Loic Magne, Anas Awadalla, Guanzhi Wang. It's a multi-institutional collaboration. Check out Guanzhi's technical deep dive thread and repo links below!

Introducing NitroGen, an open foundation model for generalist gaming agents! https://t.co/TEv0G8QqaV ✅ Open access to the largest action-labeled gameplay videos with 40K hours across 1,000+ game titles ✅ Universal simulator for measuring cross-game generalization ✅ Open gaming foundation model trained to play 1,000+ games with large-scale behavior cloning 🧵

Website: https://t.co/SgkyODmmzR Paper: https://t.co/RYCZqku6iM Code: https://t.co/BWRwakEhkW Pretrained model weights: https://t.co/Ab0rBf4wHc Action dataset: https://t.co/ZoLHrpaCjY Deep dive from @guanzhi_wang https://t.co/a1VyQEMjxd

+1 more

Do you want to run coding agents safely, without damaging to your filesystem? 📁 Last week, we published a blog post and a demo showing exactly how to do this with @claudeai and AgentFS by @tursodatabase. After strong community interest, we’ve now shipped support for @OpenAI Codex as well 🚢 How it works: 💻 Launch the filesystem MCP server 🆕 Open a new demo session 🚀 Start coding with Codex Supporting Codex unlocks a big advantage: developers can use any OpenAI-compatible provider, including @ollama and @huggingface Inference API. This means more flexibility and safer experimentation, all without compromising your local environment. Let us know what you build with it! 👩‍💻 Find the code on GitHub: https://t.co/lCeaFCYHve 📚 Read the blog: https://t.co/IiCW8Bo0NZ

+1 more

We are going to unlock the future of our skies and ensure America — not China leads the world in this technology. 🇺🇸 https://t.co/joLXf9s2yW

This is the story of how reading “The Selfish Gene” when I was around 15 changed my career decades later. It’s a terrific book. But beyond its substance, it changed my view of what science can be. It showed me that there are simple but profound ideas waiting to be discovered. I’d thought of the frontier of science as necessarily esoteric, but the book proved otherwise. Richard Dawkins’s writing also showed me that it’s possible to explain novel and profound ideas in a way that even a child (me) could understand them. When I grew up and became a researcher, I never stopped thinking about this. I began to gravitate toward the simplest questions within my areas of expertise, rather than the hardest, contrary to the norm in science. And I taught myself how to communicate my ideas to as broad an audience as possible. Unfortunately, the peer review process heavily penalizes this approach, because the value system prioritizes abstruseness, when ideally it should be the opposite. But no matter — I found that simpler ideas, when they do get published, are much more widely read, which made it all worthwhile. Besides, pushing to make ideas as simple and as simply communicated as possible often made them *better ideas*, more robust and widely applicable than initially anticipated. Aspirations should be balanced with an awareness of one’s limitations. Not everyone can be as successful as Dawkins; I realized that I couldn’t count on my ideas being so powerful that they would spread on their own (fittingly, the term “meme” was coined in The Selfish Gene!) So I’ve tried to put as much effort into spreading ideas as I do into generating and explaining them. That’s a topic I’ve written about here before and probably will again.

🤗 MiMo-V2-Flash is live on @huggingface , supported by Novita Test it yourself👇 https://t.co/tXJb5QwVuv

🤗 MiMo-V2-Flash is live on @huggingface , supported by Novita Test it yourself👇 https://t.co/tXJb5QwVuv

i will cherish it forever - thanks @huggingface / @pollenrobotics! https://t.co/jX3rqroIwe

i will cherish it forever - thanks @huggingface / @pollenrobotics! https://t.co/jX3rqroIwe

+2 more

At @huggingface, how we created doc chatbot system using open weights LLMs & open source software stack: 1. chunk & embed docs markdowns using open weights embedding models (BAAI/bge-base-en-v1.5, google/embeddinggemma-300m, we usually use https://t.co/zWv8ZjEiMd to find the hottest models), use @meilisearch (https://t.co/kU8wkaIuyV) as our semantic search db 2. https://t.co/0GeysYbFab that acts as gateway to semantically search hf docs (among other things). https://t.co/DFnlv8W55z by @evalstate 3. add https://t.co/0GeysYbFab to https://t.co/WXiwAO0VCK, which uses latest open weights LLMs such as zai-org/GLM-4.6 (https://t.co/74frBh1wpa) through HF Inference Providers (https://t.co/6j1SC6JyeU) that runs your inference through @FireworksAI_HQ, @togethercompute, @GroqInc, @novita_labs

+5 more

On one hand, the vast majority of people surveyed get this wrong, on the other hand, how many people can explain how other technologies they use work? https://t.co/iwojGqqrIA https://t.co/kN1Hvny1NJ

📣#AWS & @atomico are proud to partner on the State of European Tech report 2025. https://t.co/LBR5guNjBv 🔮The future will be written by those who can move fast & think long-term. Discover opportunities for the #startup community & beyond, with in-depth findings from the region.

New perks for Enterprise organizations to engage with the AI community on @huggingface 📢 show your company with your profile activity 💼 show you're hiring on your profile I show you how in the video https://t.co/d3kpwK5eoF

BREAKING: Assistant U.S. Attorney Joe Thompson exposes the true level of fraud in Minnesota. "Every day we look under a rock and find a new $50 million fraud scheme." "[Minnesota] is an outlier in a bad way. You don't see fraud on this scale in other states." "Traditional Medicare/Medicaid fraud is people overbilling. Upcoding. Instead of billing for a service with a relatively modest reimbursement, you claim you're providing a higher dollar value service. You're maybe making 20% more - padding your bills." "What I think is unique to Minnesota is that's not what we're looking at. What we're seeing is programs that are just entirely fraudulent. These aren't companies that are providing some services but overbilling Medicare and Medicaid, these are companies that are providing NO services. They're essentially shell companies created to defraud the program. That's unique to Minnesota."

California bill requires teachers to undergo LGBTQ propaganda training, which promotes "preferred pronouns," classifies Heterosexism as "bigotry," and hides children's gender identity from parents. Gavin Newscum wants to trans your kids. https://t.co/RdfWelX4ou

Two Haitians in Massachusetts have been criminally charged after being caught trafficking nearly $7 million worth of food stamp benefits and sold donated food product intended for the nonprofit ‘Feed My Starving Children’ https://t.co/teaccTtI6E

Elon Musk: Never trust a cynic, because most people are kind of medium good ❤️ “Never trust a cynic. The reason is, that cynics excuse their own bad behavior by saying everyone does it, because they're cynical. So I always be, it's a red flag if someone's a cynic, a true cynic. But this is an important point that I think people listening should bear in mind. If somebody is cynical, meaning that they see bad behavior in everyone, it's easy for them to excuse their own bad behavior by saying, well, everyone does it, but it's not true. Most people are kind of medium good.”

🚨HOLY SH*T - THE UK IS GOING FULL POLICE STATE Not only are they getting rid of Jury trials They plan to allow magistrates to sentence people for up to two years without appeal ‼️ Last year 40% of appeals from mags courts were upheld. LABOUR WANT TO LOCK UP INNOCENT PEOPLE https://t.co/P3jve4CHGp

The Northern Virginia Justice System are Liberal Illegal Alien Sympathizers who put anarchists above public safety. Marvin Morales-Ortez - an illegal alien from El Salvador - who committed homicide a day after being released from jail by yet another liberal judge. He was in jail for brandishing a gun and assaulting and injuring someone. Since 2020 he has been charged with SEVEN crimes in Fairfax County - one of which was First Degree Murder. This is just outside of DC - an illegal alien enclave protected by woke Liberal judges and prosecutors. https://t.co/xhpuFaKAwt

🚨🇩🇪 GERMANY’S “WELCOME CULTURE” COLLAPSES: MAJORITY NOW WANT IMMIGRATION HALTED & DEPORTATIONS RAMPED UP A new poll just dropped, and it’s a political earthquake: a majority of Germans now support an immigration moratorium and mass deportations. The poll shows that 81% of Germans state that illegal immigration was “far too high” or “rather too high” over the last 10 years, and only 5% view the levels as appropriate. Another 75% say illegal migration has been bad for Germany. Yes, in Germany… the same country that once opened its doors to over a million refugees in 2015, which was promoted by Angela Merkel, her CDU, and the left. However, after years of failed integration, rising crime headlines, housing shortages, and mounting public frustration, the mood has flipped hard. Nowadays, 53% of Germans favor a total moratorium on all new immigration and immigration overall, as well as the return of a large number of migrants to their homelands, otherwise known by many on the right as “remigration” or mass deportation. It’s official… the once-celebrated “welcome culture” has turned into: enough is enough. Berlin’s political class can’t keep ignoring this shift, because the public clearly isn’t. Source: @zerohedge, Remix News, YouGov

Elon Musk painting the ultimate post-scarcity future: AI + robots → unlimited production → universal high income → money becomes meaningless → we design dream homes on demand, travel in tunnel EVs or electric VTOLs, and live for fulfillment, not survival. As he mentioned it is straight out of Iain M. Banks’ Culture series, I would read them all to understand the future. But start here: Consider Phlebas: an epic space opera showing the Culture at war, full of pirates, megastructures, and a glimpse of what true abundance looks like (even if seen through an enemy’s eyes). This is the future we’re building and will arrive at.

Test drive a Cybertruck, It will blow your mind Elon Musk https://t.co/yumgZiPwE7

Morgan Freeman on Elon Musk: "I'm a huge fan of Elon Musk. I think he's got the most incredibly forward-thinking ideas about where we can go technologically." https://t.co/7KxAj923Jp

The only thing standing between you and outrageous success is hard work. https://t.co/MKHaZLOncj

The only thing standing between you and outrageous success is hard work. https://t.co/MKHaZLOncj

Deep Reinforcement Learning for Multi-Agent Coordination Multi-agent reinforcement learning doesn't scale. The default approach to coordinating robot teams remains explicit communication or centralized training. This is inefficient. On the other hand, social insects solve this problem effortlessly. Ant colonies coordinate thousands of agents through stigmergy: indirect communication via environmental traces. This new research introduces S-MADRL, a framework where robot teams coordinate through virtual pheromones instead of direct communication. Agents leave digital traces in a shared virtual map as they move through the environment. Other agents sense these traces within their local field of view and incorporate them into their decision-making. The environment itself becomes the communication medium. The researchers combined this stigmergic approach with curriculum learning, training agents sequentially rather than simultaneously to address non-stationarity issues. The emergent behaviors of the system mirror biological strategies. Agents self-organized into bucket-brigade patterns with bidirectional flow. Some agents spontaneously became idle during high congestion, exactly like worker ants in crowded tunnels. No one programmed these behaviors. They emerged from the stigmergic signals alone. Paper: https://t.co/CqiJrpctY6 Learn to build effective AI Agents in our academy: https://t.co/Y5kVy5iKiQ

Designing reward functions for RL agents is kind of broken. The default approach remains manual engineering: domain experts iteratively craft reward signals through trial-and-error. This requires significant expertise, takes enormous human effort, and often fails when task complexity increases. But what if agents could discover their own optimal reward functions? This new research introduces a bilevel optimization framework that automatically discovers optimal reward functions for embodied RL agents through regret minimization. The optimal reward function can be defined as one that minimizes the gap between the learned policy and the true optimal policy. No expert demonstrations needed. No human feedback required. How it works: Two optimization levels run simultaneously. The lower level trains the RL agent to maximize rewards as usual. The upper level continuously updates the reward function itself, guided by a meta-gradient that minimizes policy regret. The reward function learns to assign high values to critical states like success or failure, while providing dense feedback throughout the state space. The framework works across both value-based agents (DQN) and policy-based agents (PPO, SAC, TD3) without task-specific tuning. In data center energy management, all RL agents using discovered rewards achieved energy reductions exceeding 60%, compared to 21-52% for baseline RL. In UAV trajectory tracking, the approach enabled PPO agents to successfully complete tasks where hand-designed rewards failed entirely. In sparse-reward OpenAI benchmarks, agents using discovered rewards outperformed baselines in both convergence speed and final performance. The discovered reward functions also reveal interpretable structure: they automatically identify critical states and encode latent relationships between states and rewards that match physics-based reward designs, despite having no explicit mathematical model. Paper: https://t.co/W9fRH6sbDq Learn to build effective AI Agents in our academy: https://t.co/JBU5beIoD0

Brilliant post on prompt caching! This is one of the most effective and underutilized techniques for reducing LLM usage costs. https://t.co/nByCrQv5ez

History LLMs are models trained exclusively on pre-1913 texts. Beyond the research applications, here is why this is an exciting effort for me: 1. Studying historical discourse without modern bias. These models capture what was "thinkable, predictable, or sayable" at specific moments in history. Unlike prompting modern LLMs to roleplay, these models genuinely don't know about future events because that information literally isn't in their training data. Lots of applications there. 2. Understanding historical predictions and assumptions. Researchers can explore what contemporaries expected would happen versus what actually occurred. This is useful for studying economic forecasts, political analysis, and social expectations from past eras. 3. Analyzing language and concept evolution. Track how terminology, ideas, and discourse patterns changed over time. The specific cutoff dates (1913, 1929, 1933, 1939, 1946) align with major historical inflection points (pre-WWI, Great Depression, WWII start, post-war). Heck, this might even be useful where you are using sub-agents for history-related tasks or expertise. 4. Detecting anachronisms and large-scale textual analysis. Useful for historians, writers, and filmmakers to verify period-accurate language and concepts. These models can flag modern assumptions that wouldn't exist in historical contexts. This also enables exploration of massive historical corpora in ways traditional archival research cannot. It acts as a "compressed representation" of the discourse from each era. Thoughts?