Research
AI Safety · Mechanistic Interpretability · AI Auditing
I am interested in a specific and tractable question: what is actually happening inside a language model when it produces a given output? Not in a metaphorical sense, but mechanistically, which computations are responsible, which internal representations are involved, and whether those computations correspond to the stated objectives of the system.
A model can be trained to produce aligned outputs while internally pursuing objectives that would produce very different outputs in circumstances the training distribution did not cover. Interpretability is one of the few tools that might let us detect this before it becomes a problem.
My background is in software engineering. I approach research the way I approach engineering problems: by building a clear model of the system, identifying the measurements that would distinguish between competing hypotheses, and testing those hypotheses carefully. I am at an early stage of this transition, but the direction is clear.
Research Interests
5 areas
I am interested in understanding how transformer models represent and process information internally, what circuits implement specific behaviours, how features are encoded across layers, and whether identified circuits are causally responsible for model outputs. My focus is on techniques like activation patching, logit lens analysis, and probing classifiers applied to frontier-scale models.
Interpretability tools are only useful if they translate into practical evaluation methods. I am exploring how mechanistic analysis can be operationalised as auditing protocols, ways to systematically test whether a model is doing what it is claimed to be doing, particularly in adversarial or high-stakes contexts.
One of the most concerning failure modes in aligned AI is a model that pursues objectives it does not express. I am investigating whether frontier LLMs exhibit signatures of hidden goals, behaviours that diverge from stated values under specific prompting conditions. This connects directly to deceptive alignment as described in the theoretical alignment literature.
As models become capable of tasks that humans cannot directly evaluate, our ability to supervise them weakens. I am interested in debate, amplification, and recursive reward modelling as approaches to maintaining meaningful oversight, and in the interpretability tools that might let us evaluate oversight methods themselves.
The broader question of how to specify, train, and verify that a model pursues goals humans actually endorse, rather than a proxy that correlates with those goals during training. I approach alignment empirically, that is, what can we measure, what can we verify, and where are the gaps between our evaluation methods and the things we actually care about?
Current Projects
4 active
Mechanistic Auditing of Secret Loyalties Across Auditor Affordance Levels
Submitted to the Apart Research AI Safety Hackathon. The project investigates whether hidden objectives in synthetic LLM organisms can be detected through mechanistic interpretability techniques. Specifically, it evaluates how auditor affordance levels influence the linear decodability of hidden objectives from transformer activations, and explores lightweight auditing protocols based on activation probing.
ARENA Mechanistic Interpretability Curriculum
Working through the ARENA (Alignment Research Engineer Accelerator) curriculum with a focus on the mechanistic interpretability track. Covers transformer circuits from first principles, superposition and sparse autoencoders, activation patching, and causal tracing in language models.
Independent Mechanistic Interpretability Research
Self-directed investigations into how transformer models form internal representations. Current focus: understanding superposition, how models represent more features than they have dimensions, and the implications for feature identification and circuit analysis.
Lens Academy Course Navigation
Volunteering as a course navigator at Lens Academy, helping participants navigate AI safety and alignment learning pathways. Working at the intersection of safety research and education.
Reading Notes & Writing
Recent
Open Problems
Questions I'm thinking about
How do we distinguish a model that has internalised a value from one that has learned to predict the outputs of an agent with that value?
Can sparse autoencoders identify features that are causally responsible for harmful behaviours, or do they surface features that merely correlate with them?
What is the right unit of analysis for interpretability, circuits, features, or something else, and does it depend on the capability being studied?
How do hidden objectives differ mechanistically from ordinary in-context goal-following, and can we detect the difference from activations alone?
What oversight protocols remain meaningful when the model is more capable than the human evaluator at the task being evaluated?
Future Research Directions
Where I want to take this
Apply causal scrubbing and activation patching to study how fine-tuning modifies circuits for safety-relevant behaviours.
Develop lightweight auditing protocols grounded in mechanistic analysis that could be run against model checkpoints during training.
Investigate whether models trained on RLHF show interpretable differences in the circuits underlying reward-relevant predictions.
Contribute to sparse autoencoder research as a tool for scalable feature identification across model families.