Today's Training Data episode takes us BTS on the infrastructure challenges required to do large RL runs at scale, featuring @ellev3n11 (Composer Lead at @cursor_ai) and @dzhulgakov (Co-Founder at @FireworksAI_HQ). The Cursor team trained Composer 2 on Fireworks by starting with a strong base model (Kimi 2.5) and performing large-scale mid-training on code tokens and web data to learn common patterns and libraries, followed by a large-scale Reinforcement Learning run to learn how to navigate the Cursor harness, call tools, and write correct code. Today's episode dives into the systems and infrastructure challenges of making that large RL run happening, and there were many (!!), from numerical mismatch to global distribution to synchronizing rollouts across asynchronous pipelines to keeping track of expert activation across runs and more. Extremely nerdy in-the-weeds challenges that Federico and Dima were delighted to nerd out on together :) Beyond RL infra, we also discussed Online vs Simulated rollouts, self-summarization for long-horizon agents, environment design ("the most powerful RL environment is the product itself"), and other technical nuggets. PS: We filmed this episode before the SpaceX news, while the Cursor team was still compute-constrained. While Cursor now has *all* the flops, the takeaways and hurdles crossed ring true for any serious application-level company that is racing to post-train their own models. I believe that more serious application companies will go the way of Cursor and post-train their own models. 00:00 Introduction 00:53 Why Cursor Trained Composer 2 04:55 Specialization vs Bitter Lesson 06:16 Composer 2 Training Recipe 16:32 Scaling RL Infrastructure Globally 23:32 Floating Point Drift 25:11 MoE Sensitivity Explained 26:25 Router Replay Fix 27:19 Real Time RL Loop 31:49 Long Horizon Agents 34:29 Why RL Everywhere 37:34 LLM as Judge Rewards 39:14 RL in Hard Domains 40:13 Build Your Own Environments 44:34 Closing Thoughts

Introducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.

Cursor beating every model that exists on CursorBench

Composer 2 is now available in Cursor.

Composer 2 is now available in Cursor.

Composer 2 is now available in Cursor.

Composer 2 is now available in Cursor.

Today we introduce humans&, a human-centric frontier AI lab. We believe AI can be reimagined, centering around people and their relationships with each other. At its best, AI should serve as a deeper connective tissue that strengthens organizations and communities

We are at NeurIPS and excited about compute-optimal RL with value functions! Come stop by our poster on Friday :)

Happy to announce our neurips’25 paper, real world RL of active perception behaviors! I am pretty excited about this project - I learned that real world robot RL is actually quite straightforward. Details below:

I'll be at Neurips this week! On Friday afternoon (4:30-7:30), we're presenting our work on training value functions compute-optimally. value-scaling.github.io/ With @_oleh @zhiyuan_zhou_ @mic_nau @pabbeel @svlevine @aviral_kumar2

Claude Opus 4.5's score on SWE-bench is wild. I like how Anthropic has focused on coding from the beginning. They haven’t released any image or video models. All in the most economically valuable area. Good strategy.