Research Note: Filtering Subversion-Relevant Information From Pretraining Data Is Feasible
Kyle O’Brien1, Spencer Kitts2, Cameron Tice1 and Alek Westover2
1 Geodesic Research; 2 Redwood Research
AI Safety Research · Cambridge, United Kingdom
Geodesic
Research
The shortest path to impact
About Us
We build robust initialisations for alignment—base models shaped through pre- and midtraining to hold up under capabilities reinforcement learning.
Geodesic Research is a UK-based technical AI safety organisation focused on compute-intensive alignment research. Our seminal work on alignment pretraining showed that you can bake alignment priors into base models, and the broader field is now converging on the approach.
Long-horizon capabilities reinforcement learning is emerging as a critical and underexplored threat to alignment—degrading alignment properties across evals and selecting for behaviours like metagaming, sycophancy, and reward hacking. Our agenda is to design midtraining and early post-training interventions that create initialisations where alignment persists through the rest of training.
Research Taste
01
Conceptually Simple
We focus on conceptually simple, data-centric interventions—document mixes, filtering, declarative midtraining—that benefit from scale and slot into existing training pipelines without bespoke infrastructure.
03
Frontier Impact
Our target audience is model training teams at frontier labs. We design interventions that can be profiled, packaged, and handed off—taking the shortest path to advising on their production training stacks.
Research & Other Writings
Reducing the Resource Gap Between Lab and External Safety Researchers
Alexandra Narin, Kyle O’Brien and Puria Radmard
Geodesic Research
Inoculation Midtraining with Learned Neologisms
Kyle O’Brien1, Edward James Young1, Puria Radmard1, Nathalie Kirch1, Cameron Tice1, Tomek Korbak2 and David Demitri Africa3
1 Geodesic Research; 2 OpenAI; 3 UK AI Security Institute
Why study proto-training gaming as an adversarial alignment failure mode?
Puria Radmard, Edward James Young and Cameron Tice
Geodesic Research
Why study alignment interventions on pre-RL checkpoints?
Edward James Young, Puria Radmard and Cameron Tice
Geodesic Research
Announcing Geodesic Research
Puria Radmard, Cameron Tice, Alexandra Narin, Edward James Young and Kyle O’Brien
Geodesic Research
Alignment Pretraining: AI Discourse Causes Self-Fulfilling (Mis)alignment
Cameron Tice1, Puria Radmard1,2, Samuel Ratnam3, Andy Kim4, David Demitri Africa5 and Kyle O’Brien1
1 Geodesic Research; 2 University of Cambridge; 3 University of Oxford; 4 Independent; 5 UK AI Security Institute
The Team
Founded in Cambridge, UK
Co-Founder & Executive Director
Marshall Scholar at the University of Cambridge, where he completed his MPhil on automated research with LLMs for computational psychiatry. A former Research Manager for the ERA:AI fellowship.
Co-Founder & Technical Director
Former ERA:AI fellow and University of Cambridge PhD student. Previously a machine learning engineer at raft.ai and a private equity quantitative strategist at Goldman Sachs.
Alexandra Narin
Chief Operating Officer
Founding Member of Geodesic & Cofounder of UKAIF. Previously, an Experimental Neuroscience researcher at UCL and the Head of Grants for a Biotech company.
Edward Young
Founding Member of Technical Staff
Former researcher on AISI's Safeguards team and ERA:AI Fellow. Completed a Computational Neuroscience PhD at the University of Cambridge.
Founding Member of Technical Staff
Leads the alignment pretraining research agenda and has developed strong relationships with UK AI Security Institute through previous research on Deep Ignorance. Previously at EleutherAI and Microsoft.
Member of Technical Staff
PhD student in computer science at Imperial College London and King's College London, researching mechanistic interpretability and robustness in LLMs. Previously a MATS Research Scholar, LASR fellow, and ERA:AI fellow.
Nathaniel Mitrani Hadida
Member of Technical Staff
PhD student at the University of Cambridge. Researched model organisms of scheming as a MATS fellow and alignment faking as an ERA:AI fellow. Previously worked on multilingual and multimodal capability evaluations at Qualcomm AI Research.
Ryan Lundqvist
Member of Technical Staff (Work Trial)
Former Pivotal Fellow and SPAR Fellow, researching situational awareness and metagaming. Previously, worked on causal inference models as an engineer at Meta and graduated from Georgia Tech, where he was in the GT AISI.
Ben Maltbie
Member of Technical Staff (Work Trial)
Former Pivotal Fellow and recently completed a master’s in EECS at MIT, where he was part of MIT AI Alignment (MAIA). Previously a software engineer at Amazon.
Join Us
Help build the future
of AI safety
Geodesic is hiring Members of Technical Staff.
We're looking for technical staff with experience across the ML and alignment research stack: multi-GPU / HPC training and evals experience, deep familiarity with data-centric alignment methods, and an insatiable desire to improve the outcomes of developing superintelligence.
If this sounds like you, please apply.