Alberti: Multi-Turn RL Enables Aggressive Code Optimization Over Single-Turn Models
Silas Alberti · DeepWiki: The GitHub Encyclopedia · May 21, 2025 · at 30:19
Silas Alberti of Cognition explains the technical findings from training an open-source model converting Python code to high-performance CUDA kernels using reinforcement learning.
0:00 / 0:21exact quote · 21.6s
▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark
“Basically the single-turn model that was just trained on, like, getting the best result after one turn. It would basically be a little bit, like, too careful, because it couldn't risk writing, like, non-compiling code, whereas, like, the multi-turn model would maybe do, like, more aggressive optimization, but then, like, there's, like, one compiler or one bug, and it has, like, three more turns to fix it, so it was kind of nice.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Silas Alberti
Alberti: AI App Companies Must Encode Product Needs Into RL Feedback
“I think that's almost how I view like the future of application layer companies. Cause I mean, yeah, you see like the different, the labs are also now creating these like RL platforms and you can soon like customize models with RL on your personal, like on you…”
Alberti: Small numbers of manually curated evals beat large eval sets
“I think usually like small numbers of very high quality evals are the way to go and like we curate them manually and make sure they're really good.”
Alberti: Wiki pages are a better abstraction than pure RAG for codebase search
“If you just do like pure, like rack on like such a big base code files, it'll just be like pretty bad at a certain point, you know, on a single code base. Sure. I can see it, but Tens of thousands of code bases. It's tougher. But I think actually the wiki page…”
Alberti: DeepWiki compute spend may break $1M soon
“So we think we should be now at like the high, like tens of thousands of repos. So I guess you can like do the rough map. Like we might be approaching quite significant numbers of compute spend here. I mean maybe we'll like break the million soon.”
Alberti: Folder structure graphs are poor representations of codebase architecture
“There are some, the simplest one is a folder structure graph, but unfortunately, it's, like, actually a pretty bad one because people sometimes just, you know, like, put all the components in this folder and all the servers in this folder.”
Alberti: Cognition built its own in-house vector database
“This culture of like building everything in-house. It's also kind of funny, like we, there's like a rack component to this. And I've been actually, I personally advocate, I can just like use a vector database and make life easy. But then there's a really stron…”