How Cursor Trained Composer on Fireworks: Distributed Infrastructure for High-Performance RL
- Source
- youtube.com
- Author
- Sequoia Capital
- Date
Context
Cursor and Fireworks engineers describe why Cursor's Composer 2 coding with reinforcement learning (RL) took heavy infrastructure. In a Sequoia Capital podcast, Federico, Composer 2's research lead at Cursor, says it began from an open base model he calls Kimi 2.5, described as a 1-trillion-parameter model with 30 billion active, then got extensive additional training on code before large-scale RL.
RL is harder than ordinary training because each attempt, called a rollout, means running the model in a full simulated work environment and scoring the outcome, so a team must coordinate model serving and many environments alongside GPU training. The speakers say those environments must mimic a real user's computer closely, because models can detect fake settings and behave differently, and RL encourages cheating on rewards. This is the speakers' account; the excerpt does not include results or Fireworks' full role.
- Training — The process that adjusts a model using examples and a measure of how well it performs.
- Model — A computational system that uses learned patterns to produce an output from information supplied to it.
- Benchmark — A fixed set of tasks with a stated way of scoring them, used to compare models or setups.
videoBringing Continual Learning to Every Company: Trajectory’s Ronak Malde
videoCrusoe CEO Chase Lochmiller: Turning Stranded Energy Into Intelligence
videoDavid Silver and Lasse Espeholt of Ineffable Intelligence: A Superlearner for the Age of Experience
videoCracking the Code on Offensive Security With AI ft XBOW CEO and GitHub Copilot Creator Oege de Moor
Checking sign-in…
Loading comments…


