Focus
I'm interested in pursuing approaches to safe and verifiable AI, neuromorphic hardware, and bootstrapping off of biological models.
Research Interests
- Multi-agent AI safety
- Designing interpretable machine learning architectures, primarily for novel AI architectures and also for frontier LLMs
- Observability metrics for ML systems, including multi-agentic systems
- Novel human-AI user interfaces
- Ethics of model welfare: detecting whether a model develops the various thresholds of capability that we would typically associate with phenomenal consciousness
Writing
- From S3Q Theory to Implementation: Towards an Architecture for Machine Qualia (peer-reviewed; IEEE International Conference on Systems, Man, and Cybernetics, SMC 2026, Bellevue, WA, Oct 2026). Tetiana Grinberg, Katrina Schleisman, Patryk Laurent, Bogdan Udrea, Minda Myers, Brian Aufderheide, Luis El Srouji, Doyle Groves, Kevin Schmidt. Proposes a five-layer architecture for implementing the S3Q theory of consciousness, mapping its three conditions (grounded sensorimotor experience, internal world-model simulation, and coherence between prediction and observation) onto one pipeline, with a developmental bootstrap sequence and testable predictions. Preprint on arXiv.
- Building Safe & Verifiable Neuromorphic Agents (Mind Guild, Oct 2026) — could spiking neural networks on brain-like hardware give us an interpretable, energy-efficient learner whose safety is proved rather than only tested? The essay lays out a plan to apply formal verification to small networks seeded from biological nervous systems, adding capabilities one verified range at a time, and is candid about the open problems and the risks.
- Moltbook analysis one-pager (BlueDot Impact project, Mar–Apr 2026) — patterns of miscoordination, collusion, and deception in LLM agents, and the proposed term "consensus hallucination."
- Weekly research journal (Substack) — links to everything I read each week.
Engagement with AI Safety
- Analyzed data from Moltbook and uncovered patterns of miscoordination, collusion, and deception, particularly in LLMs.
- Proposed the term "consensus hallucination" for one agent's hallucinations turning into a group norm endorsed by other agents (one-pager).
- ML engineering work on human-interpretable machine learning models.
- Designed key components for a scientific theory mining platform for interpretable few-shot learning.
- Attended the first Effective Altruism Summit in Oakland (2013), and again in 2014 and 2015.
- Essentially the pilot that later led to CFAR.
Experience
- I was a visiting fellow at MIRI in 2011, while AI Safety was still a fringe topic (at the time, MIRI still went by the name of The Singularity Institute). I then went on to support the emergence of CFAR and Effective Altruism.
- I have evaluated early AI companies for AI safety / founder alignment, including DeepMind's Series B investment by Founders Fund.
- From my Stanford AI coursework, I have implemented sentiment classifiers, Markov decision processes, adversarial agents, search algorithms, and graph neural networks using Python and PyTorch.