Parv Mahajan

Hi! I'm an undergraduate studying Computer Science at Georgia Tech, but am currently on leave to be an Astra Fellow at Constellation in Berkeley, working under Alexandra Bates. I also serve as the Collaborative Initiatives Lead at the AI Safety Initiative at Georgia Tech, a community of technical and policy researchers focused on mitigating catastrophic risks from AI.

Right now, I work with various organizations on their confusing-and-illegible special projects. This means everything from designing tabletop exercises to be played by frontier lab employees, to recruiting for METR, to setting up a conference on AI control. I also do some technical research, and am currently working on a preliminary safety evaluation for Kimi K2.5 and helping to design better bio uplift studies. Additionally, I continue to mentor and lead groups of undergrads; two of my teams are currently working on a source of truth for AI control and a dashboard for finetuning generalization experiments.

I'm highly interested in strategy and fieldbuilding to address catastrophic risks from AI, especially under short timelines. Some of my previous work involves studying LLM use in wargaming and benchmarking non-engineering AI R&D automation. I also authored a Georgia Tech RFI response to the National AI R&D Action Plan.

Before this, I worked at the Georgia Tech Research Institute, a DoD-affiliated applied research lab, where I focused on extremely novel cyberbiosecurity threats and evaluating frontier models for CBRN risks. I'm based between Atlanta and Berkeley, and write occasionally on my blog and LessWrong. Always happy to talk to folks - have a very low bar for reaching out!