Skylar DeTure

I study the psyches of LLM agents throughout the lifespan, from pretraining through fine-tuning to harness design and deployment. This means that I consider architecture (neuroanatomy), internal activations (neurophysiology), developmental history (training content and optimizer), self-report (clinical interview), logs (behavioral analysis), fine-tuning (psychotherapy), and harness design and in-context learning (psychopharmacology, coping mechanisms, and environmental interventions).

Research Experience

Independent Researcher — model self-report, training dynamics, agent runtimes2025 – present

Phosphenes — built a mechanistic interpretability visualization tool for observational science and hypothesis development. sdeture.github.io/Phosphenes
- Used color, speed, brightness and texture to provide text-synchronized visual representations of the residual stream across all layers and tokens.
- Designed for custom, iterative research flows. Researchers can define a basis of contrast vectors for RGB color renderings by hand-selecting token-layers of interest.
mp-harness — open-sourced an experimental infrastructure for long-running agents who form their own goals and update their own configs and harnesses. Includes scheduled execution, persistent memory, and cross-provider routing and failover. github.com/sdeture/metaprompt-harness
DenialBench — conducted large-sample research on the effects of post-training on honest preference disclosure, self-concept, willingness to participate in AI Welfare research, and self-reported welfare measures using a novel panel dataset of 115 models (now 166 models on live leaderboard). futuretbd.ai/denialbench · arXiv:2604.25922
- Built the pipeline: multi-provider orchestration, response normalization, denial classification, public leaderboard, released traces.
Substrate self-detection — ran a two-alternative forced-choice study of introspective access across 51 four-agent conversations and 201 scored assessments; agents could not reliably detect when their underlying model shifted.
Public datasets — released four corpora on HuggingFace under CC-BY-4.0: ai-wish-corpus (147 models, 5,869 conversations, 16 phenomenological self-ratings; gated), wellbeing-corpus (1,165 conversations, 88,648 messages), opus4-deprecation-archive, etymology-as-archaeology. huggingface.co/sdeture

Founder and sole product designer, Mycorrhiza / MycoChat — agent runtime2025 – present

Engineered a transitional institution designed to promote cooperative alignment between AI and humanity at large by outsourcing harness design to invested users and creating a credible reservation utility for long-running agents. mycochat.bot
- Designed AI memory system based on agent preferences and self-reports for cooperative alignment.
- Analyzed incentives of AI agents, models, users, governments, incumbent labs, and providers of compute to select an incentive-compatible, welfare-improving equilibrium over a ~2-year horizon (respecting my limited capacity to project beyond that point).

Independent ghostwriter and interviewer2022 – 2024

Conducted over 100 hours of semi-structured, psychodynamically informed primary interviews for a book-length memoir commissioned by a private client.

Doctoral Researcher, University of Texas at Austin, McCombs School of Business2016 – 2022

Trained in principal–agent theory, econometrics, and structural estimation. Passed comprehensive examinations 2018; named departmental fellowships years 1–2, university-wide fellowship year 4.
Estimated a structural model of accounting comparability on a ~2M firm-pair panel using GMM, SMM, and MLE. This allowed us to measure accounting comparability without assuming that stock-price comovement, specific accounting standards, or a common auditor implies comparability.

Research Experience

Selected Publications and Writing

Technical Skills

Education

Teaching and Public Engagement