"The state of things is evidently far from simple" - William James
Latest highlighted read:
Things I worked on
Fieldbuilding, leading the policy efforts in AI to ensure a positive future.
learn more →
Worked with the Institute for AI Policy and Strategy on robotics threat modeling.
Currently published: a reading list geared for a policy audience that should take <45 mins, and commentary on the FCC's restrictions on new foreign advanced robotics.
Working with Stepan Shabalin to apply SAEs for mech interp on VLAs.
A paper analyzing US legislation trends around deepfakes in the 2020-2025 time period, produced as my final artifact for INTA8000 under Prof. Margaret Kosal.
read here →
A published Google Docs extension for selective word counting. 2K+ downloads!
website + download →Sources I recommend
Mostly policy, love their weekly briefings!
visit source →More technical, great analysis of model cards.
visit source →More anthropological/social commentary shaped.
visit source →Exactly what you'd assume from the title.
visit source →Great books
An absurdist comedy all about self-forgiveness with outstandingly well-written female characters. Probably my favorite work of literature ever, the dialogue is both histerical and poignant.
The prime example of "esperpento" a Spanish play genre that can be understood as being a tragedy with strong social critique. Beautiful but soul-crushing, I was in a funk for a few days.
Ted Chiang is a favorite between sci-fi fans and this short story along with "Hell is the Absence of God" are my favorite.
Fantastic long-form essay-argument, made me grateful for the randomness inherent in human talent and encouraged my first public Substack
It's up there with Fahrenheit 451, but I found the concept of SOMA more troubling. Weirdly, I did not enjoy 1984 as much.
Very interesting reflections on the role of the law. Reading it was the first time I encountered the argument that suing and leveraging the law as the civilized alternative to direct violent acts is merely a change in the form of violence. The one suing simply defers their desire for retribution to the law. You can find the PDF here.
Hey, I'm Rocio! B.S./M.S. Comp Sci @ Georgia Tech, minor in Law, Science and Technology. think of the c as a th and you're good to go.
I am interested in understanding what good AI policy and governance actually looks like.
I co-direct the AI Safety Initiative at Georgia Tech, leading field-building efforts to get the very best working on preventing harms from transformative AI.
Read my Substack
Give me anonymous feedback
Email me: rocio.perales (at) gatech.edu
What I'm up to
- IAPS summer fellowship, worked on physical AI threat modeling, read my explainer on IAPS' Attack Surface on the FCC restrictions on new foreign advanced robotics, here
- Co-directing AISI, read about us here
- Interpretability research on VLAs in the PAIR Lab
- Reviewed the Spanish translation of METR's OpenAI/Hugging Face incident investigation report
If you're at Tech and want to get involved with AISI, come to our next event and feel free to reach out
Education
Skills
Professional & Research Experience
- Leading interpretability SAE-based work on VLAs (π0.5) targeting singular feature ablation to test feature transferability from base to finetuned models, advised by Stepan Shabalin and Prof. Animesh Garg
- Assisted in ICRA paper on Progress Reward Model training achieving a 10-15% performance jump on precise tasks
- Synthesized a 269-page Great American AI Act discussion draft into a 2-page decision brief for the Policy Team
- Authored a public explainer and policy recommendations on the FCC's proposed ban on foreign-made advanced robotics, securing review from 5 policy experts through direct outreach
- Conducted internal robotics threat modeling for the Frontier Security team, including BOTEC calculations on drone attacks and interviews with 6 stakeholders across industry, military, academia, and think tanks
- Trained YOLO models on AWS GPU instances and integrated autoencoders for powerline anomaly detection
- Google Docs extension allowing for tailored word counts with 3K+ downloads
Leadership & Campus Involvement
- Selected by Provost Raheem Beyah to chair subcommittee for Georgia Tech's institutional AI strategy
- Designed and deployed a campus-wide AI usage survey collecting 1K+ responses without monetary incentives
- Lead a 15-person exec organization with $30K+ of external funding and set overall strategy
- Created a Policy Fellowship curriculum run as a course seminar with Georgia Tech's OMSCS program
- Led deepfake governance research under Dr. Margaret Kosal
- Led implementation on the multi-robot pathfinding system (C++/ROS2, real-time decision-making under uncertainty); placed 2nd in the technical challenge and 4th overall at RoboCup 2025 Intl