Activity
30K followers
-
Robert Nishihara reposted thisRobert Nishihara reposted thisAs more teams scale open source AI on secure, governed infrastructure, Ray has become critical across the full AI lifecycle, from data curation to production serving and reinforcement learning. Ahead of PyTorch Conference North America 2026 in San Jose, we put together a guide highlighting key Ray-focused sessions on the schedule. Featured speakers include: Anyscale / Databricks / University of California, Berkeley: Ion Stoica Anyscale: Eric Tang, Sumanth R Hegde, Joshua Lee, Mengjin Yan Google: Ankita Luthra, Kotturu Trinadh, Jago Macleod LinkedIn: Tao Huang, Tommy Li Pinterest: Gaurav Arora, Shunyao Li, Eric W. Uber: Ke Chen, Peng Zhang, Xandra Zhu University of Minnesota: Arun Sharma Explore engineering takeaways from these technical sessions where you will learn about unified AI orchestration with Kubernetes, elastic training stacks, and scalable RL. Read the full guide and map out your schedule: https://lnkd.in/e373G-Hu Register for PyTorch Conference North America: https://hubs.la/Q04tBgv_0 #PyTorchCon
-
Robert Nishihara reposted thisRobert Nishihara reposted thisReflecting on Anyscale Ray Summit 2026 + vLLM Conference, where my team and I had the incredible opportunity to present our LLM serving platform at Chase, built on Ray + vLLM. We covered our work on custom speculative decoding, request routing, inference profiling, high availability, and zero-downtime serving. Also lots of discussion centered on where LLM inference, distributed systems, and RL post-training are heading.. especially as post-training itself becomes an increasingly complex systems problem. RL post-training is also an area we’re actively researching now! #vLLM #Ray #LLMInference #DistributedSystems #ReinforcementLearning
-
Robert Nishihara shared thisIon gave a great talk about gaps in agentic software engineering at Ray Summit. https://lnkd.in/gMq-rw8rRobert Nishihara shared thisOur team at Berkeley has spent six months synthesizing Just-in-Time Systems for complex applications. We've extracted the learnings from the failures we've seen into a framework that's been guiding our ongoing research. Why can automatically generated and verified code still fail in production, and what causes reward hacking and hallucinations? We answer these and more in our recent paper, summarized in this article. Joint work with Alex Krentsel, Shubham Agarwal, Mert Cemri, Shu Liu, Sidharth Sankhe, Ziming Mao, Matei Zaharia.
-
Robert Nishihara reposted thisRobert Nishihara reposted thisOur team at Berkeley has spent six months synthesizing Just-in-Time Systems for complex applications. We've extracted the learnings from the failures we've seen into a framework that's been guiding our ongoing research. Why can automatically generated and verified code still fail in production, and what causes reward hacking and hallucinations? We answer these and more in our recent paper, summarized in this article. Joint work with Alex Krentsel, Shubham Agarwal, Mert Cemri, Shu Liu, Sidharth Sankhe, Ziming Mao, Matei Zaharia.
-
Robert Nishihara reposted thisRobert Nishihara reposted thisThat's a wrap on the vLLM Conference at Ray Summit. Lines out the door for talks about an inference engine—that says everything about how much this community shows up. Thank you to Anyscale for co-hosting with us, and we're already looking forward to the next event. Every session recording is in the comments below:
-
Robert Nishihara reposted thisRobert Nishihara reposted thisIn his recent talk at Anyscale’s Ray Summit, Bedrock’s Vincent Gonguet covered how the company was able to get to commercial deployments of fully autonomous excavators in only two years, and ways they use Ray to power the stack: heterogeneous CPU and GPU pipelines, mining ambiguous multimodal field data, and a closed-loop simulator that models everything down to the dirt. Watch the full talk in the link in the comments.
-
Robert Nishihara shared thisI watched a bunch of the talks at Ray Summit live and this one by Priunsh Syen and Tyler Titsworth at Lila Sciences about the details of their AI research platform was extremely impressive. https://lnkd.in/eJxyTCwpInside Lila's AI Research Platform | Lila Sciences | Ray Summit 2026Inside Lila's AI Research Platform | Lila Sciences | Ray Summit 2026
-
Robert Nishihara reposted thisRobert Nishihara reposted thisHad a great time speaking at Ray Summit 2026 about The Synthetic Data Flywheel at Pinterest. We shared how we’re using Ray Data and vLLM to build large-scale synthetic data pipelines for post-training — combining GPU inference, tools, and LLM judges into scalable post-training workflows. Ray Data has been a great fit for these kinds of heterogeneous workloads, especially when dealing with stages that have very different scaling and concurrency characteristics. This work was truly a team effort. Huge thanks to Yash Upadhyay , Raphael Poulain, Tianjian Huang, Zhenyu Tan, Shubham Gupta, Shunyao Li, Luyan Wu, Korhan Citlak, Dave Chen, AI Training team at Pinterest! A special thank-you to Karthik A.ial thank-you to Karthik A. and Bo Liu for their leadership and support. Really enjoyed sharing what we’ve learned with the Ray community. Thanks to everyone who joined the session, to Eric W. for co-presenting this with me, and to Anyscale for putting together a great Ray Summit! #RaySummit #Ray #LLM #PostTraining
-
Robert Nishihara reposted thisRobert Nishihara reposted thisRay Summit 2026 is a wrap. More than 2,000 people came to San Francisco on August 24-26 to ask the question: what does it actually take to run reinforcement learning at scale? Lila Sciences is turning a 15,000-square-foot robotics lab into a verifier for scientific RL, with results spanning CAR-T development, mRNA design, and alloy screening. Torc Robotics consolidated five training systems onto one Ray cluster and is now running one billion autonomous trucking simulation miles per week. Bryan Catanzaro from NVIDIA walked through how Ray coordinates 3,000-plus GPUs for Nemotron RL training, with a 13 percent throughput improvement from topology-aware scheduling alone. Periodic Labs, Bedrock Robotics, Spotify, Capital One, Microsoft AI, Recursion, and more brought the same depth to the breakouts. Read the full recap. https://lnkd.in/gwrGc-zH
-
Robert Nishihara liked thisRobert Nishihara liked thisComfy API is live Until now, shipping a ComfyUI workflow to production meant rebuilding its environment somewhere else. Renting GPUs, reinstalling every custom node and model, untangling Python dependencies, and writing your own scaling logic. Comfy API handles that for you: → Builder reads your workflow JSON, finds the models and custom nodes it needs, and helps resolve dependency conflicts → Each Build pins the ComfyUI version, nodes, models, and Python deps together → Releases are immutable, so the environment you test is the one you deploy → Endpoints autoscale, down to zero when idle or with warm workers when latency matters → GPU time is billed by the second on RTX PRO 6000, H100, H200, or B200 Your graph stays yours, the engine stays open source, and your Builds stay portable. Click the link below to deploy your first workflow today
-
Robert Nishihara liked thisRobert Nishihara liked thisAs more teams scale open source AI on secure, governed infrastructure, Ray has become critical across the full AI lifecycle, from data curation to production serving and reinforcement learning. Ahead of PyTorch Conference North America 2026 in San Jose, we put together a guide highlighting key Ray-focused sessions on the schedule. Featured speakers include: Anyscale / Databricks / University of California, Berkeley: Ion Stoica Anyscale: Eric Tang, Sumanth R Hegde, Joshua Lee, Mengjin Yan Google: Ankita Luthra, Kotturu Trinadh, Jago Macleod LinkedIn: Tao Huang, Tommy Li Pinterest: Gaurav Arora, Shunyao Li, Eric W. Uber: Ke Chen, Peng Zhang, Xandra Zhu University of Minnesota: Arun Sharma Explore engineering takeaways from these technical sessions where you will learn about unified AI orchestration with Kubernetes, elastic training stacks, and scalable RL. Read the full guide and map out your schedule: https://lnkd.in/e373G-Hu Register for PyTorch Conference North America: https://hubs.la/Q04tBgv_0 #PyTorchCon
-
Robert Nishihara reacted on thisRobert Nishihara reacted on thisReflecting on Anyscale Ray Summit 2026 + vLLM Conference, where my team and I had the incredible opportunity to present our LLM serving platform at Chase, built on Ray + vLLM. We covered our work on custom speculative decoding, request routing, inference profiling, high availability, and zero-downtime serving. Also lots of discussion centered on where LLM inference, distributed systems, and RL post-training are heading.. especially as post-training itself becomes an increasingly complex systems problem. RL post-training is also an area we’re actively researching now! #vLLM #Ray #LLMInference #DistributedSystems #ReinforcementLearning
-
Robert Nishihara liked thisRobert Nishihara liked thisIn light of the progress in mathematics, we at Edison Scientific and FutureHouse have assembled a set of Millennium Problems for Biology. They are chosen to be very hard to solve but very easy to validate in a simple laboratory environment. Any of these, if solved, would mark a major advance in biotechnology, and most of them would contribute materially towards curing disease. These are, in some sense, the “last reasonable eval” for AI in biology. This was work primarily by Michaela Hinks and myself, with contributions from many others. Short descriptions below. The full descriptions of the problems with acceptance criteria are at millenniumproblems.bio. Share more if you have ideas. If they meet our criteria, we’ll add them to our list (with attribution and permission). A few notes on exclusion/inclusion criteria: -We placed a major emphasis on ease of validation, so challenges like “solve aging” are omitted. These should all be validatable in a standard wet lab in a short period of time (days, maybe a week or two.) -We placed a major emphasis on ensuring grading would be as objective as possible, so many great basic science mysteries are also omitted, because it would be very challenging to tell when the mystery was “solved.” (E.g., “uncover the function of the Vault protein,” or “explain how the Oxytricha genome works.”) -We placed a major emphasis on ensuring the problems would be understandable to anyone with a basic biology background, so some highly technical but very important challenges were also omitted. Finally, we anticipate that some problems may require edits in response to community feedback. If you think we got them wrong, please send the feedback, and we will collate and update. Thanks to Adam Marblestone, Michael Baym, Erika DeBenedictis, and Tony Kulesa for reviews.
Experience & Education
Other similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content