Nick Meyer
I'm a Masters student at Columbia University and AI safety and alignment researcher.
I'm currently working on two research projects. In the first, supported by the Cosmos Institute, I'm exploring how LLMs (mis)understand values and how an agent-mediated world can lead to a systematic degradation of autonomy. In the other, supported by BlueDot, Juan Cadile and I are figuring out how to better conceptualize introspection in LLMs and operationalize the study of it. We just presented a preliminary version of our work at the NYU Workshop on Introspection and Self-Report in Language Models.
More generally, I've been concentrating on questions related to model psychology and character, value alignment, and (dis)empowerment. I have deep interests in moral and political philosophy, as well as the philosophy of mind, which I seek to integrate into my technical work.
I'm also a graduate TA for Soren Schlassa's course on Language Models and Philosophy and am the facilitator of Columbia's AI Alignment Club Advanced Technical Fellowship.
I spend a lot of time thinking, but at the end of the day, life to me is all about feeling. Like many people in this field, I’m drawn to working on things that both excite me intellectually and bring about a feeling of muditā*.
I'm grateful for the people, art, and experiences that make me feel—the warmth of vulnerability, the reverberation of rolling basslines, the first sip of black coffee† on a slow morning.
Poke around and get to know me :)
Introspection Fine-Tuning
Replicated IFT, developed an improved objective function, and fitted jacobian lenses on Llama 3.2 and Qwen3.
Harmlessness Finetuning
Finetuning Llama 3.2 1B with SFT and DPO + LoRA for improved harmlessness.
Cambridge Alignment Bootcamp
My work through the ARENA curriculum in interpretability and alignment as part of CBAI's CAMBRIA program.
Corporate Orthogonality
Structural misalignment and the limit of technical approaches.
Dimensionality Reduction
A comparison of PCA and linear autoencoders on MNIST.
What are we building for?
Artificial intelligence, alignment, and building a future worth wanting.
Input-Attribution Methods Cannot Justify AI Decisions
Mechanistic interpretability as the path forward towards achieving a duty of justification.
We Have Yet to Cut Off the Head of the King
A misdiagnosis of power and Foucault's challenge to domination theory.
