Robert Nishihara
@robertnishihara
Followers
8K
Following
4K
Media
130
Statuses
2K
Co-founder @anyscalecompute. Co-creator of @raydistributed. Previously PhD ML at Berkeley.
Joined March 2009
Beyond pre-training, here's how I imagine most learning will work. 1. AI models / systems will maintain large collections of retrievable knowledge. This will include facts like "the capital of California is Sacramento" and tactics like "when playing Monopoly, buy a bunch of.
We're missing techniques for "training-time reasoning." Right now there's a lot of progress on inference-time reasoning, which is incredibly cool (I use o3 all the time). If I think about how I learn stuff, e.g., when reading a technical paper, it's very compute intensive. Most.
1
5
45
Everyone talks about how voice mode (once polished) will be a major UX unlock for AI, which is correct. An equally important frontier, which no one has touched yet, is AI group chats. Lots of hard product challenges to solve there, but it'll be hard to imagine AI without it once.
2
1
11
I started reading this thread and then got distracted trying to solve the math problem. It's a great problem and very enjoyable to think about. I highly encourage you to get out a sheet of paper, draw some triangles, and take a crack at it.
2/N We evaluated our models on the 2025 IMO problems under the same rules as human contestants: two 4.5 hour exam sessions, no tools or internet, reading the official problem statements, and writing natural language proofs.
0
0
8
Reinforcement learning is a big investment area for us at @anyscalecompute, and we're hiring actively for RL! If you're interested in building systems & algorithms for RL, message me.
Congratulations to my brilliant co-founder Philipp Moritz (@pcmoritz) and the legendary John Schulman, Sergey Levine, Pieter Abbeel, and Michael Jordan on their Test-of-Time Honorable Mention at ICML 2025 today!. For creating TRPO. This was done during the previous wave of
0
0
9
RT @ashugarg: Huge congrats to @pcmoritz, co-founder of @anyscalecompute for the Test-of-Time Honorable Mention at #ICML2025.
0
1
0
In large part due to Philipp's work on TRPO, reinforcement learning was one of the original motivating use cases that led us to build @raydistributed. You can see how we framed it in our early Ray paper (on page 1).
Congratulations to my brilliant co-founder Philipp Moritz (@pcmoritz) and the legendary John Schulman, Sergey Levine, Pieter Abbeel, and Michael Jordan on their Test-of-Time Honorable Mention at ICML 2025 today!. For creating TRPO. This was done during the previous wave of
0
6
17
Congratulations to my brilliant co-founder Philipp Moritz (@pcmoritz) and the legendary John Schulman, Sergey Levine, Pieter Abbeel, and Michael Jordan on their Test-of-Time Honorable Mention at ICML 2025 today!. For creating TRPO. This was done during the previous wave of
1
10
132
RT @turingcom: What a night at the Turing × @FoundationCap × @anyscalecompute happy hour!.@ashugarg, @robertnishihara & @jonsidd broke do….
0
6
0
*Lessons* close the loop between reasoning and learning.
@karpathy This concept of lessons also provides a natural way to close the loop between "reasoning" and "learning". If you do a ton of reasoning to solve a problem (at inference time), the output of that reasoning is some "lessons" (facts, insights, strategies, etc) that can then be.
0
0
2
I recently asked a friend how building a search engine for AI agents differs from building a search engine for humans (the answer wasn't obvious to me at first glance). An insight that stood out to me is that a search engine for agents can be far more controllable. A human is.
13
8
133
Join us Wednesday evening!.
🚀 Heading to #ICML2025 in Vancouver? Join us for drinks, snacks, & hot takes with Anyscale + @turingcom. - Kimi K2.- context engineering.- lesson-based learning.- the fossil fuel of AI. 📅 Wed Jul 16, 7-10 PM downtown. Spots limited — RSVP 👉
0
1
4
RT @bigdata: How WeChat (Tencent) Built #AI Infrastructure Supporting Millions of Nodes with Ray and Kubernetes.(@raydistributed @kubernete….
anyscale.com
See how the Tencent Weixin team implemented Ray and Kubernetes to build ultra-large-scale distributed systems with Ray.
0
4
0
Kimi K2 looks awesome. This blog post talks a little bit about how they use @raydistributed for data processing.
alibabacloud.com
This article introduces how Moonshot AI uses Alibaba Cloud's solutions to enhance data preprocessing for its large model, Kimi, focusing on stability, resource elasticity, and efficient management.
🚀 Hello, Kimi K2! Open-Source Agentic Model!.🔹 1T total / 32B active MoE model.🔹 SOTA on SWE Bench Verified, Tau2 & AceBench among open models.🔹Strong in coding and agentic tasks.🐤 Multimodal & thought-mode not supported for now. With Kimi K2, advanced agentic intelligence
0
7
34
This is the biggest Ray event all year and the primary conference for AI infra folks.
The Ray Summit CfP is closing on July 14!. We'll be back in San Francisco from Nov 3-5 for Ray Summit and want to showcase your work. Whether it's scaling smoother, building GenAI workflows, or creating complex ML systems - if you've built it with Ray, we want to hear about it.
1
1
5
Speak at Ray Summit (Nov 3-5 in SF)!.
have you ever wanted to speak at an AI conference? :) 📢 . we've opened our CfP for Ray Summit 2025 San Francisco, and we'd love to have you submit a talk on anything and everything ML + AI + distributed computing! . first time speakers are welcome 🫶.link below 👇.
0
0
4