Remember the llm.c repro of the GPT-2 (124M) training run? It took 45 min on 8xH100. Since then, Keller Jordan (and by now many others) have iterated on that extensively in the new modded-nanogpt repo that achieves the same result, now in only 5 min! Love this repo 👏 600 LOC

Congrats, Max! I'm so grateful for the chance to work with and learn from Max over the past few months. His combination of brain power, excitement, creativity, and kindness is off the charts and elevates everyone around him.

Depth cameras are dead

I'm really fascinated by this dataset from the AI poetry survey paper. Here's another visualization I just made. Survey respondents were shown one of these 10 poems, and either told that they were authored by AI, human, or not told anything.

Zero-shot extrapolation for out-of-distribution (OOD) chemical property prediction is an important step towards high-performance materials discovery. Check out our spotlight at the #NeurIPS AI for Accelerated Materials Design Workshop! openreview.net/pdf?id=HkfnueE…

Learning new tasks with imitation learning often requires hundreds of demos. Check out our #NeurIPS paper in which we learn new tasks from few demos by inverting them into the latent space of a generative model pre-trained on a set of base tasks. avivne.github.io/ftl-igm/

Overcoming the lack of reliability of Behavior cloning (BC) with reactive reinforcement learning. Action-chunking is a two-edged sword- it's critical for BC to work, but it also limits how adaptive the robot is to disturbances and corner cases. Learn more:

HOT 🔥 fastest, most precise, and most capable hand control setup ever... Less than $450 and fully open-source 🤯 by Hugging Face, Rob Knight, Martino Russi This tendon-driven technology will disrupt robotics! Retweet to accelerate its democratization 🚀 A thread 🧵

🧩 Why do task vectors exist in pretrained LLMs? Our new research uncovers how transformers form internal abstractions and the mechanisms behind in-context learning(ICL).

HNY! Lately I took a crack at implementing the pi0 model from Physical Intelligence PaliGemma VLM (2.3B fine-tuned) + 0.3B "action expert" MoE + block attention Flow matching w/ action chunking Strong eval on Simpler w/ 75ms inference github.com/allenzren/open… ckpts available! 👇(1/6)

Presenting Unsupervised Actuator Nets (UANs) that push the limits of agile whole-body control without the need for reward shaping! ⚡️ UANs reduce the sim2real gap in robot's motors removing the need for reward design to bridge the sim2real gap. ⚡️ A pre-trained whole-body

> benchmarking on video games > everyone is talking about RL > OpenAI has a robotics team I am no longer entirely sure what year it is

Attending #ICLR2025 next week! I will be presenting Diffusion Policy Policy Optimization (DPPO) at the Friday morning poster session with Lars Ankile diffusion-ppo.github.io I also joined Physical Intelligence lately. Love to chat about what we've been up to at Pi!

[1/3] Thrilled to be presenting our work DART tomorrow morning at ICRA! Even more excited to announce that our app is now publicly available on the App Store 🎉 Try it yourself! apps.apple.com/us/app/dart-ro… The best part? The MuJoCo simulator that powers DART now runs fully locally

1/4 We recreated a $200k teleoperation setup in VR for just ~$2k. Now we can collect more dextrous manipulation data in a single day (40 hrs/day) than any existing open dataset has ever collected.