Berkeley Artificial Intelligence Research

What are the odds that an LLM can write a highly optimized GPU kernel—with no team of GPU experts manually tuning it? What are the odds that you can take the optimization knowledge accumulated in CUDA kernels and transfer it to a very different architecture such as Apple Silicon? Ask most GPU engineers whether an LLM can do this on its own, and the answer would be: basically zero. Together with University of California, Berkeley Sky Lab, we have been building on K-Search—its AI-driven evolutionary kernel-search framework—and extending it to leverage optimization knowledge accumulated in CUDA kernels when targeting entirely different architectures and programming environments. On Apple Silicon, our evolved attention kernel reaches approximately 97% of the performance of Apple’s state-of-the-art native MLX attention kernel. Our evolved Mamba SSM kernel also delivers approximately 20× faster prefill performance than the community MLX-LM implementation. These kernels were generated without manual tuning by GPU experts. What made this work was not simply asking an LLM to write GPU code, but using K-Search to apply fine-grained optimizations and iteratively evolve kernels on the target hardware. Apple Silicon is where we started, and we plan to support more hardware architectures soon. Stay tuned! The full story is now live on the BAIR—Berkeley Artificial Intelligence Research—blog. Link in my first comment. #GPU #CUDA #AppleSilicon #MLX #KernelOptimization #AI

Until today, I had never posted about any of my research articles on LinkedIn. For the first time, however, I wanted to share this new paper, published in Nature #Cities, the result of a collaboration between University of California, Berkeley, Google and Stanford University. This publication represents the culmination of more than 15 years of research on routing that began at the Institute of Transportation Studies Berkeley California PATH - ITS Research Center at UC Berkeley, in collaboration with Nokia, NAVTEQ, Caltrans - California Department of Transportation, U.S. Department of Transportation and many others, and was later pursued at Uber and now Google. This work presents the first large-scale deployment of a routing system that actively distributes traffic across alternative routes, reducing congestion-related CO₂ emissions while people are already on the road. Thank you to Andrew Tomkins, Neha Arora, Théophile Cabannes, Abdul Rahman Kreidieh, Shoshana Vasserman, Marc Nunkesser, Prem Ramaswami, Eray Turkel, PhD, Kevin Chen, and Haizheng Zhang for this marathon together! Looking back, it is remarkable to see how far the field has come since the days of #MobileMillennium, the first traffic application deployed in North America on #Rim and #Symbian smartphones to crowdsource #GPS data for real-time traffic estimation, together with Quinn Jacobson, John Paul Shen, Ken Tracton, Dan Work, Ryan Herring, Aude Hofleitner, Timothy Hunter, Jeff Ban and many others. Since then, years of research have built on the fundamental insight behind John Nash's equilibrium: when everyone independently chooses the fastest route, that route often ceases to be the fastest. Google post: https://lnkd.in/g4PwPQq5 Nature Cities article: https://lnkd.in/gCAfx6cD

📌 How well can today’s VLMs actually perform on your robotics task? Check out RoboVista at RSS 2026! 🤖👀 University of California, Berkeley Google DeepMind, Princeton University Many robotics researchers have tried giving a VLM an image from their own task, asking a task-specific question, and seeing what it can understand. As VLMs become more capable, they are increasingly being integrated into robotic pipelines for perception, reasoning, planning, and decision-making. In the future, robots will work across agriculture 🌱, factories 🏭, hospitals 🩺, homes 🏠, roads 🚗, and unstructured environments 🌲. This raises an important question: How well do today’s VLMs understand robots at work across these diverse domains? We’re excited to present RoboVista at #RSS2026 a benchmark for evaluating VLMs across diverse real-world robot applications. RoboVista includes: • 🧠 474 expert-annotated questions • 🤖 39 robotics task types • 🌍 6 domains: agriculture, industry, surgery, domestic robotics, autonomous driving, and open robot datasets • 🔍 Questions spanning perception, planning, motion reasoning, and failure recovery 🚩 Key findings 🧠 More reasoning is not always better: Chain-of-Thought can hurt low-level perception through overthinking, even when it helps with multi-step planning. 👀 Perception remains the main bottleneck: larger models reduce misidentification errors, but spatial and geometry-aware reasoning remain far from solved. Huge thanks for Kaiyuan Chen co-lead the project and all the great collaborators Ziyang Chen, Simeon Oluwafunmilore Adebola, Yixuan Huang, Zehan Ma, Tianshuang Qiu,Wentao Yuan , Dhruv Shah, Pannag Sanketi, Ken Goldberg! #Robotics #VisionLanguageModels #EmbodiedAI #RobotLearning #PhysicalAI Have you noticed Chain-of-Thought causing "overthinking" errors in your own VLM experiments? Let me know below! 👇 Explore the benchmark, leaderboard, dataset, and more findings in the comments.

Congratulations to Berkeley Artificial Intelligence Research alumnus Deepak Pathak has been awarded the PAMI Young Researcher in Computer Vision Award. This top award for young researchers in computer vision is given to two recipients yearly. Deepak Pathak is the seventh Berkeley Artificial Intelligence Research affiliated awardee in the last seven years. Other awardees include Angjoo Kanazawa (2024), Judy Hoffman (2023), Bharath Hariharan (2022), Georgia Gkioxari (2021), Phillip Isola (2021), and Jon Barron (2020). https://t.co/PuVbUpp6oo

Are you applying to grad school this fall? The Equal Access to Application Assistance (EAAA) program for UC Berkeley Electrical Engineering & Computer Sciences (EECS) is now accepting applications! Any PhD applicant to UC Berkeley Electrical Engineering & Computer Sciences (EECS) can submit their application for feedback by Oct. 5th 11:59PM PST. EAAA is a student-led program designed to help more grad applicants get access to application assistance. A current or recent grad student in Berkeley EECS will provide feedback on application statements, resumes, and other application materials prior to the application deadline. Please see https://lnkd.in/geEuDwan for more information and send any questions to eaaa@berkeley.edu!

“Data will solve robotics and automation: True or false?”: A debate. Held at IEEE International Conference on Robotics and Automation (ICRA) #ICRA2025 and featuring BAIR faculty Ken Goldberg and BAIR alumnus Animesh Garg.

New Science Robotics editorial by BAIR Faculty Ken Goldberg: Good old-fashioned engineering can close the 100,000-year “data gap” in robotics.