Then I discovered Modal. Suddenly my agent could spin up compute on-demand with minimal code change—1 or 5 GPU, A100 or H100, whatever it needed, exactly when it needed it. https://t.co/fp8RfDoHO8

This led me to the concept I now call agent-oriented compute — a way of designing systems where AI agents can request exactly the compute they need, when they need it.

I think we're at this weird inflection point. We're building increasingly capable AI agents, but we're forcing them to use compute infrastructure designed for humans submitting computational jobs. (It’s like teaching AI agents to control the browser by clicking buttons.)

An AI agent that can dynamically request and manage compute would unlock massive potential — faster iteration, large-scale experimentation, and parallel exploration that no human-driven workflow could match.

This year, I’ve been developing an AI agent for scientific experimentation — one that can design and execute experiments to verify hypotheses. While building it, I ran into one major challenge: how to give AI agents efficient access to compute resources to do ML research.

Last night I was at @modal ’s Series B celebration party— super fun night. I think it might be interesting to share a bit about my own experience with Modal.

If you’re interested in agentic intelligence—LLMs that don’t just think but act—I go into the nuts and bolts of what it takes to make this work in practice. https://t.co/BCLidNFZrC

In the blog, I cover: - How RL for LLM-based agents differs from traditional RL for LLM. - The critical system challenges when scaling agentic RL. - Emerging solutions top labs and companies are using

TL;DR: The frontier of AI is moving from simple-response generation to solving complex, multi-step problems through agents. Previous RL frameworks for LLMs aren’t built for this—they struggle with the heavy, diverse resource demands that agents need.

While working on this, I notice the underlying RL systems must evolve to support these new capabilities. So, I wrote a blog post to capture my thoughts and lessons learned. “When LLMs Grow Hands and Feet, How to Design our Agentic RL Systems?”

Lately I’ve been building AI agents for scientific research. In addition to build better agent scaffold, to make AI agents truly useful, LLMs need to do more than just think—they need to use tools, run code, and interact with complex environments. That’s why we need Agentic RL.

Over the past five years, we have witnessed the rapid expansion of AI from a technology confined to specialized domains within large enterprise (recommendation, surveillance) to a tool that is increasingly integrated into the everyday life of individuals (ChatGPT, Cursor..)

EXP-Bench: Can AI Conduct AI Research Experiments? "EXP-Bench challenges AI agents to formulate hypotheses, design and implement experimental procedures, execute them, and analyze results." "EXP-Bench curated 461 AI research tasks from 51 top-tier AI research papers."

We're passionate about open science and invite you to try Curie and even contribute to making it better for everyone! Check out our post: https://t.co/uIWQYffK1F

For example, Curie can generate highly performant models, achieving a 0.99 AUC (top 1% performance) for a melanoma (cancer) detection task.

Our goal is to empower researchers like them to rapidly test hypotheses and extract deep insights from their data. Curie automates the aforementioned complex ML pipeline – taking the tedious yet critical work.

needed to select the right algorithms, tune hyperparameters, or interpret model outputs, we knew we had to help. That's why we're so excited to introduce the new AutoML feature in Curie 🔬, our AI research experimentation co-scientist designed to make ML more accessible!

After personally seeing many researchers in fields like biology, materials science, and chemistry struggle to apply machine learning to their valuable domain datasets to accelerate scientific discovery and gain deeper insights, often due to the lack of specialized ML knowledge

If you’re curious, feel free to check it out! Blog, Demo, Code ⬇️

From ideation during winter break to paper submission took a bit over a month, and then another month to open source everything. Somehow, every PhD project just keeps getting more fun—now I kind of don’t want to graduate. 😆