Evan Ryan Gunter

Email: evgunter at gmail dot com

CV, last updated 2/25.

• Geometry of loss landscapes
• Dependence on specific optimizers of the loss landscape's local geometric properties during training, on which I mentored a MARS project (paper)
• Understanding to what extent different training runs lead to qualitatively different solutions; generalization of the finding that, in deep linear nets, all local minima are global to nets with nonlinearities such as ReLUs
• Understanding how LLMs work, and application of these insights to mechanistic interpretability
• How well SAEs represent true underlying features
• How LLMs perform computations on their internal representations of features
• How important attention is at scale compared to MLP layers

• Adversarial examples
• Whether adversarial examples are preventable or a necessary consequence of some aspect of how ML models work, like superposition
• Generalization of adversarial examples between models with different architectures or models trained on different datasets
• Theoretical minimum sizes of models with certain capabilites; putting bounds on greatest possible hardware overhang
• The limits of out-of-distribution generalization; especially, whether LLMs can gain capabilities qualitatively beyond those demonstrated by humans in their training data

• Self-refining alignment strategies such as Constitutional AI, Superalignment, and multimodal model self-supervision
• Non-alignment strategies for safe AI in the short term, especially leveraging the low agency of frontier models

• Moral patienthood of AI systems, and connections between goals and affective states
• Anthropic reasoning, on which I wrote a thesis with advisor Chip Sebens
  • Consequences of different formalizations of anthropic reasoning
  • Application to the many-worlds interpretation of quantum mechanics
  • Application to multiverses governed by different laws
  • Addressing mathematical issues in infinite universes
• The hard problem of consciousness and Russellian monism, on which I did a small research project with advisor Frederick Eberhardt and collaborator Alex Denko
• Philosophy of fundamental physics
• Ontology of physical laws
• Implications of the many-worlds interpretation of quantum mechanics

• Quantifying stability of non-power-seeking in artificial agents
• NGD converges to less degenerate solutions than SGD
• My GitHub

I have recently transitioned to researching AI safety. Previously I was a research engineer at Granica working on data compression. I graduated from Caltech in 2019, with B.S. degrees in mathematics (advisor Nets Katz), computer science (advisor Chris Umans), and philosophy (advisor Chip Sebens).