News for hackers. Stuff that matters.

silasalberti

silasalberti on Hacker News

Re:SWE-Grep and SWE-Grep-Mini: RL for Fast Multi-Turn Context Retrieval on Thursday October 16, 2025 @03:00PM (#45609337)
hey I'm from the SWE-grep team - feel free to ask me any questions :)

Show HN: Blockdiff – How we built our own file format for VM disk snapshots

Posted by silasalberti on Monday June 23, 2025 @02:44PM
from the show-hn dept.

https://cognition.ai/blog/blockdiff

Usually, I’m a researcher working on areas like RL for coding agents – but one day I became annoyed by our slow VM startup times. So I took the plunge into systems engineering and built the first version of our VM hypervisor called otterlink. It now powers both our research & all of Devin production workloads.Compared to EC2, otterlink was able to bring down VM startup times by about 10x....

( Read More... | 1 comment | 8 points | on HN )

Show HN: Kevin-32B – how to do multi-turn RL on writing CUDA kernels

Posted by silasalberti on Tuesday May 06, 2025 @03:48PM
from the show-hn dept.

https://cognition.ai/blog/kevin-32b

Hey – we just published a blog post about Kevin-32B = K(ernel D)evin.It's to our knowledge the first open-source model that's RL-trained on CUDA kernels. Our goal was to demonstrate multi-turn RL using GRPO. We used 180 Python->CUDA conversion tasks from the KernelBench dataset.The results were surprisingly strong! We were able to outperform top reasoning model like o3 & o4-mini.We're...

( Read More... | no comments | 2 points | on HN )
Re:DeepWiki – Which repo would you like to understand? on Friday April 25, 2025 @06:59PM (#43799286)
Currently it uses only the codebase itself. However we‘ve been thinking about adding other source like docs. Can’t promise any timelines yet
Re:DeepWiki – Which repo would you like to understand? on Friday April 25, 2025 @06:56PM (#43799262)
This is entirely custom built but we‘re also fans of Mintlify. We use them for docs.devin.ai
Re:DeepWiki – Which repo would you like to understand? on Friday April 25, 2025 @06:54PM (#43799251)
When working with LLMs you can never exclude hallucinations entirely but we‘ve been carefully tuning the system over time and found it to be pretty high signal! The reason we display the code snippets is to make it easy to double check with the source
Re:DeepWiki – Which repo would you like to understand? on Friday April 25, 2025 @01:37PM (#43796431)
Hey I'm Silas, I worked on this. Lmk if you have any questions!