- Joined
- Jun 2025
- How they joined
- Directly
- Before Cognition
- Stanford University Department of Computer ScienceML Compiler Researcher
- Now
- Flappy Airplanes
- Pietro Marsella is a co-author of the Kevin-32B blog post, listed with an asterisk indicating equal contribution.May 6, 2025
- Pietro Marsella is affiliated with Stanford University and Cognition AI.May 6, 2025
- Pietro Marsella co-authored work on Kevin-32B, a model trained with multi-turn reinforcement learning for writing CUDA kernels.May 6, 2025
- Pietro Marsella's co-authors on the Kevin-32B work are Carlo Baronio, Ben Pan, and Silas Alberti.May 6, 2025
- Pietro Marsella co-authored the Kevin-32B model, which achieves 65% average correctness across the dataset and solves 89% of it.May 6, 2025
- Pietro Marsella worked on kevin-32b, a model trained with multi-turn reinforcement learning to write fast CUDA kernels, with a blog post on cognition.ai and a paper on arXiv.2025
- Pietro Marsella worked on swe-grep, a model trained with reinforcement learning to do fast (sub 5s) codebase-wide agentic context retrieval, with a blog post on cognition.ai and a Hacker News link.2025
- Pietro Marsella has a project called "kevin-32b" described as "model go brrr".
- Pietro Marsella has a project called "swe-grep" described as "agentic codebase search".
- Pietro Marsella previously worked at Cognition on reinforcement learning for coding models.
- Pietro Marsella co-authored work using KernelBench, a dataset of 250 PyTorch-based classic deep learning tasks.May 6, 2025
- Pietro Marsella gave an excellent performance in the individual competition of the provincial stage of the Olimpiadi di Matematica, representing the province of Milan.Jun 1, 2023
RL for code generation
Machine Learning Compiler Research for the FlexFlow project.
Software developer for optical payload for ION satellites, Technology Transfer & Prototyping Division
Software developer of Alcor, a space-grade air purifier
To express full training runs in a single kernel, we had to build general control-flow primitives. This lets us efficiently run weirder, more dynamic, fine-grained workloads than any other framework (we can run DOOM!). Come join us if you want to build the future of GPU systems!
banger release as always!!
happy to share what we've been working on! making agentic search fast while not compromising on quality is key to a good user experience
Test-Time inference is cool! But also a a bit of a shoehorn: models are usually trained to generate a correct answer first try. This naturally steers them away from trying out riskier responses, even when allowed multiple attempts. Enter multi-turn training! See the 🧵for more.
